FullText PDF

Title
Author(s)
Kinematics, dynamics and control of high precision parallel

manipulators
Cheung, Wing-fung, Jacob.; .
Citation
Issued Date
URL
Rights
2007
http://hdl.handle.net/10722/51504
The author retains all proprietary rights, (such as patent rights)

and the right to use in future works.
Abstracts of thesis entitled
Kinematics, Dynamics and Control of High Precision Parallel

Manipulators
Submitted by
Cheung Wing Fung Jacob

for the degree of Doctor of Philosophy
at The University of Hong Kong
in March 2007
End-point positioning accuracy, fast point-to-point motion travel and stable operation
are essential in high-precision motion mechanisms for industrial applications such as
semiconductor packaging. In this thesis, a high-precision 2-DOF planar parallel
manipulator and a 4-DOF spatial parallel manipulator are developed to advance the
motion capabilities of traditional XY table and overcome difficulties in the
kinematics design and end-point positioning accuracy of existing parallel
manipulators. Novel methods of kinematics design and optimization, dynamic
modelling and advanced control strategies are integrated in a complete design of the
proposed manipulators to achieve high-speed positioning of the end-effector to an
accuracy of micrometer unit.
Modular design approaches are developed for designing the kinematics structure of
the proposed parallel manipulators. In this design process, consideration is given to
ensuring that the kinematics design can be realized to the required tolerance using
existing mechanical construction technology. A Monte-Carlo simulation method is
developed to determine the workspace of the parallel manipulator, and a Jacobian
decoupling approach is developed to verify that the workspace of the manipulator is
free of internal singularity. Kinematics optimization is performed to suppress the
kinematics error magnification effect so that the end-effector positioning accuracy
matches the manufacturing tolerance of the linkages.
A dynamic model is essential for motion analysis, capability estimation and

controller design of the parallel manipulator. The derivation of the dynamic model is
simplified using a nested approach whereby the 4-DOF parallel manipulator is
decomposed into triangular mechanisms at the actuator level and at the end-effector
level, resulting in a high degree of decoupling in the modelling process. Recursive
Newton-Euler method is employed to determine the dynamic model. The response of
the dynamic model is simulated using sinusoidal force profiles to drive the linear
actuators of the manipulator. The results obtained from the kinematics design,
optimization and dynamic modelling are used to develop prototypes of the planar and
4-DOF parallel manipulators for the purpose of experimentation.
A traditional PID computed-torque controller is first designed for the 2-DOF planar
and the 4-DOF parallel manipulators using a model-based approach. In view of the
limitations of the computed-torque control method, a robust learning control method
is developed to enhance the motion performance by minimizing the settling time and
steady-state error of the parallel manipulators. A frequency-domain system
identification approach is used to identify the high frequency dynamics of the
manipulator. A robust control design method is employed to design a stable, fast
tracking-response feedback controller with less sensitivity to high frequency
disturbance, with the control parameters determined using genetic algorithm. A
Fourier-series based iterative learning controller is added to the feedforward path of
the controller to improve the settling time by reducing the dynamic tracking error of
the manipulator. Experimental results demonstrate that the parallel manipulators
display considerably better on motion performance in terms of positioning accuracy,
settling time and stability than traditional XY stages. This shows that the parallel
manipulators are superior alternatives to XY motion stages for high-speed, highprecision positioning.
Kinematics, Dynamics and Control of High Precision Parallel

Manipulators
by

B.Eng.(Hons.) Sheffield; M.Sc.(Eng.) HKU; C.Eng., M.I.E.E.; M.I.E.E.E.
A thesis submitted in partial fulfillment of the requirements for

the Degree of Doctor of Philosophy
at The University of Hong Kong.
March 2007
This thesis is dedicated to my parents and my wife
Declaration
I declare that this thesis represents my own work, except where due
acknowledgement is made, and that it has not been previously included in a thesis,
dissertation or report submitted to this University or to any other institutions for a
degree, diploma or other qualifications.
Signed .
Acknowledgements
First of all, I would like to express my sincere thanks to my thesis supervisor Dr. Y.S.
Hung for his supervision, interest, support, guidance, patience and encouragement
throughout my Ph.D. research. I have benefited from his strong mathematical
knowledge, analytical analysis skills and serious research attitude. It is definitely my
pleasure to have him as my supervisor.
Secondly, I would like to thank Dr. Peter C.K. Liu, Chief Technical Officer of ASM
Pacific Technology Ltd. for his offer of this opportunity to setup this research project
and collaboration with The University of Hong Kong for me to pursue my Ph.D.
study. His continuous encouragement, support on administration, technical and
financial resources allocation established a successful research project in which the
results benefit to both academic and industrial interests.
In addition, I would like to thank Dr. D.Z. Liao, Technical Manager of R&D
Department of ASM Assembly Automation Ltd. for his untiring advice and support
on motion control implementation. I would also like to thank Dr. G.P. Widdowson,
R&D Manager of ASM Assembly Automation Ltd. for his support and guidance on
linear motor technology and Mr. M.S.W. Tam, Section Manager of ASM Assembly
Automation Ltd. for his support on servo driver technology. I must also express my
appreciation of Mr. S.K. Wong, Mechanical Engineering Specialist of ASM
Assembly Automation for his guidance and support to establish the prototypes of the
planar and the four degree-of-freedom parallel manipulators from simple kinematics
drawings.
I must also express my sincere thank to my parents for their untiring support on my
daily life and patience to my bad temper during my heavy workload pressure. I must
also thanks Joanne, my wife on providing support, encouragement and care to me
and my parents during my research study.
ii
Finally, I must express my sincere thank and praise to the almighty God, the provider
of grace, talent, ability, health, love and everything.
iii
Contents
Declaration
Acknowledgements
ii
Contents
iv
List of Symbols
vii
List of Figures
xiii
List of Tables
xxi
1. Introduction
1.1. Background
1.2. Literature survey
1.3. Research Objective and Scope
1.4. Contribution
1.5. Thesis Organization
10
2. Manipulator Kinematics
13
2.1. Introduction
13
2.2. Parallel Manipulator Development
13
2.3. Manipulator Structure
17
2.4. Kinematics of the Planar Parallel Manipulator
19
2.4.1. Forward Kinematics
20
2.4.2. Inverse Kinematics
20
2.5. Kinematics of the 4-DOF Parallel Manipulator
22
2.5.1. Forward Kinematics
22
2.5.2. Inverse Kinematics
23
2.6. Summary
25
iv
3. Jacobian, Singularity and Workspace Analysis
26
3.1. Introduction
26
3.2. Derivation of Jacobian Matrix
26
3.3. Singularity Analysis
32
3.4. Determination of the Workspace Envelope
33
3.4.1. 3D Delaunay Tessellation Method
34
3.4.2. Monte-Carlo Simulation Method
35
3.5. Summary
41
4. Kinematics Optimization
43
4.1. Introduction
43
4.2. Derivation of Error Manipulation Matrix
43
4.3. Design of Objective Functions
51
4.3.1. Minimizing the Integral of EM
2
F
over the Operational
52
Workspace
4.3.2. Minimizing the Maximum of EM
2
F
over the Operational 53
Workspace
4.4. Optimization Process
54
4.5. Results
55
4.6. Summary
60
5. Dynamic Modelling of the Parallel Manipulator
62
5.1. Introduction
62
5.2. Dynamic Modelling using Recursive Newton-Euler Method
63
5.3. Dynamic Modelling of the Planar Mechanism
64
5.4. Dynamic Modelling of the Vertical Mechanism
67
5.5. Modelling of the Belt Mechanism
72
5.6. Dynamic Simulations
74
5.7. Summary
80
5.8. Appendix
81
5.8.1. Expressions for matrices in the dynamic model (5.11)
81
5.8.2. Expressions for Matrices in the dynamic model (5.20)
81
6. Prototype Design
84
6.1. Introduction
84
6.2. Mechanical Design of the Manipulator Prototype
85
6.3. Electronic Design of the Manipulator Prototype
90
6.4. Experimental Verification of the Dynamic Model
95
6.5. Summary
98
7. Control Design of the Planar Parallel Manipulator
100
7.1. Introduction
100
7.2. Computed-torque Control Design
101
7.3. Robust Control Design
105
7.4. Robust Learning Control Design
109
7.5. Experimental Results
113
7.6. Summary
130
8. Control Design of the 4-DOF Parallel Manipulator
132
8.1. Introduction
132
8.2. Computed-torque Control Design
133
8.3. Robust Control Design
135
8.4. Robust Learning Control Design
135
8.5. Experimental Results
136
8.6. Summary
163
9. Conclusion and Future Research Work
165
9.1. Conclusion
165
9.2. Future Research Work
167
List of Publications
168
Awards
168
References
170
vi
List of Symbols
OXYZ
Rectangular coordinate system for the base frame of the 4DOF parallel manipulator
d1, d3, d5, d7
Position of the linear actuators 1, 3, 5, and 7
P(x1, y1)
Point on the planar mechanism with the actuator pairs (d1, d3)
to construct the vertical T-mechanism
Q(x2, y2)
Point on the planar mechanism with the actuator pairs (d5, d7)
to construct the vertical T-mechanism
l2, l4, l6, l8, l
Length of the linkages of the planar triangular mechanisms
k10, k12, k
Length of the vertical linkages of the 4-DOF parallel

manipulator
G(x,y,z,)
Coordinate of the end-effector of the 4-DOF parallel

manipulator
Position vector of the linear actuators
q
Velocity vector of the linear actuators
Position vector of the end-effector of the 4-DOF parallel

manipulator
X
Velocity vector of the end-effector of the 4-DOF parallel

manipulator
Jacobian matrix of the 4-DOF parallel manipulator
Xp
Position vector of the small platform of the T-mechanisms
X p
Velocity vector of the small platform of the T-mechanisms
Jq
Jacobian matrix of the linear actuators and the small platforms
JH
Jacobian matrix of the small platforms of the T-mechanisms

and the end-effector
Number of planar surface sliced from the vertical axis of the

workspace of the 4-DOF parallel manipulator
zi
Position of the data point i in z-axis of the workspace
vii
Vertical distance between each planar surface of the
Zd
worksapce
xh(i), yh(i)
Coordinate of the ith data point of the workspace in the hth

planar surface used for edge detection
Ph(xh(i), yh(i))
Subset of the data points of the workspace envelope located in

the hth planar surface
gc(h)
Number of column of the grid table at the planar surface h
gw(h)
Number of row of the grid table at the planar surface h
rx(h)
Column resolution of the grid table at the planar surface h
ry(h)
Row resolution of the grid table at the planar surface h
l2, l4
Manufacturing variation of the length of linkages l2 and l4
l6, l8
Manufacturing variation of the length of linkages l6 and l8
k10, k12
Manufacturing variation of the length of the vertical linkages

k10 and k12
( x1 , y1 )
Position of the point P of the planar mechanism caused by the

length variation of the linkages l2 and l4
(x1, y1)
Position variation of (x1, y1) due to the length variation of l2

and l4
(x2, y2)
Position variation of (x2, y2) due to the length variation of l6

and l8
( x, y, z , )
End-effector position of the 4-DOF parallel manipulator

caused by the ...
(x, y, z, ) Kinematics error of the end-effector due to the variation of the

small platform position and the length of vertical linkages
Vector of the kinematics error of the end-effector
Vector of the manufacturing error of the linkages
EM
Error manipulation matrix
Singular values of the error manipulation matrix
EM
Frobenius norm of the error manipulation matrix
f (l, k )
Operational workspace of the parallel manipulator

Objective function of the product of singular values over the
operational workspace
viii
f1 ( l , k )
Objective function of the integral of EM
2
F
over the
operational workspace
f2 (l, k )
Objective function for minimizing the worst-case kinematics

error over the operational workspace
Actuating force vector of the parallel manipulator
Mass matrix of the parallel manipulator
Velocity coupling vector of the parallel manipulator
Gravitational force vector of the parallel manipulator
Fd
External force vector of the parallel manipulator
fi*
Inertia force exerted at the centre of mass of link i expressed

in frame i
i *
i
Moment exerted at the centre of mass of link i expressed in

frame i
mi
Mass of link i
Absolute linear acceleration of the centre of mass of link i
vci
expressed in frame i
i
Absolute angular velocity of link i expressed in frame i
i
Absolute angular acceleration of link i expressed in frame i
Inertia matrix of link i about its centre of mass expressed in
Ii
frame i
i1
Ri
Matrix representing the rotation from frame (i1) to frame i
i1
fi,i-1
Reaction force exerted on link i by link (i1) expressed in

frame (i1)
i1
ni,i-1
Moment exerted on link i by link (i1) expressed in frame

(i1)
ri
Vector from the origin of frame (i1) to the origin of frame i

expressed in frame i
rci
Vector from the origin of frame i to the centre of mass of link
i expressed in frame i
p , p
Interaction forces in the Y- and X-directions (respectively)

acting between the two serial limbs at the joint connecting
links 2 and 4
ix
p , p
External forces in the Y- and X-directions (due to the vertical

mechanism) acting between the two serial limbs at the joint
connecting links 2 and 4
, ,
Interaction forces about a vertical axis acting between the two

serial limbs at the joint connecting links 10 and 12
Moment about a vertical axis acting between the two serial

limbs at the joint connecting links 10 and 12
, ,
External forces about a vertical axis acting between the two

External moment about a vertical axis acting between the two

Force constant of the sinusoidal force profile
Phase angle of the sinusoidal force profile
Frequency of the sinusoidal force profile
Ui(t)
Controller output of the linear motor i
Uai(t), Ubi(t)
Two phase control voltage of the linear motor i
ei ( t )
Electrical angle determined by the position of the linear motor

i and the magnetic pole-pitch
Iai(t),Ibi(t),Ici(t) Three-phase driving currents of the linear motor i

Kc
Current gain of the servo amplifier
qd
Desired joint acceleration vector of the linear actuators
q
Actual joint acceleration vector of the linear actuators
Control input to the manipulator in computed-torque

controller
Position feedback control output of the computed-torque

controller
c ( s )
Closed-loop characteristics polynomial of the computedtorque controller
KP , KI , KD
Diagonal control gains of the computed-torque controller
Damping ratio of the joint error i
Natural frequency of the joint error i
Pi
Joint-space dynamic model of the parallel manipulator
Li
Low frequency model of the parallel manipulator
Hi
High frequency model of the parallel manipulator
Ki
System gain of the translational joint i
Viscous friction coefficient of the translational joint i
Resonant frequency of the translational joint
Anti-resonant frequency of the translational joint
Damping ratio of the resonant frequency of the translational

joint
Damping ratio of the anti-resonant frequency of the

translational joint
Gc
PID controller
FL
Lag compensator
FN
Notch filter
Position error of the translational joint
uc
Control output of the robust PID controller
Time constant of the robust PID controller
KL
Normalized gain of the lag compensator
fp
Lower corner frequency of the lag compensator
fc
Higher corner frequency of the lag compensator
Attenuation of the notch filter
High break-frequency of the notch filter
Low break-frequency of the notch filter
uf
Control output of the feedforward controller
Kv
Velocity feedforward constant
Ka
Acceleration feedforward constant
Kj
Jerk feedforward constant
Tm
Motion time for a complete point-to-point travel
Tp
Trajectory time of a point-to-to motion
Tt
Desired settling time after the end of motion
Td
Delay time between two motion cycles
xi
Ts
Sampling time of the real-time feedback controller
j (k ) , j +1 (k ) Inverse Fourier transform of the filtered spectrum of the error

signal in the cycle j and (j+1)
Gl
Learning loop gain
Forgetting factor of the learning loop
j +1 (k )
Control output before regulation at cycle (j+1)
j +1 (k )
Regulated iterative learning control output by the PD

controller at cycle (j+1)
uf
Output signal of the learning control loop
xii
List of Figures
Fig. 1.1
Block diagram of a typical robot manipulation and
control system
Fig. 2.1
Stewart platform
14
Fig. 2.2
DELTA parallel manipulator
14
Fig. 2.3
H4 parallel manipulator
14
Fig. 2.4
Manta parallel manipulator
14
Fig. 2.5
Triangular linkage
16
Fig. 2.6
H-linkage
16
Fig. 2.7
V-linkage
16
Fig. 2.8
Triangular mechanism with end-effector
16
Fig. 2.9
Triangular mechanism with end-effector and
16
parallelogram compensation mechanism

Fig. 2.10 Two triangular mechanisms arranged in orthogonal
17
configuration
Fig. 2.11 Two triangular mechanisms arranged in parallel
17
configuration
Fig. 2.12 Kinematics design of the 4-DOF parallel manipulator
18
Fig. 3.1
35
Workspace envelope of the 4-DOF parallel manipulator

determined by 3D Delaunay Tessellation method
Fig. 3.2
Grid table of mapping the x-y coordinates of the data
36
points in the manipulator workspace at the specific

planar surface
Fig. 3.3
Mapping of the grid table of the workspace data points
37
to a binary image
Fig. 3.4

observed from the origin
xiii
38
Fig. 3.5
Workspace envelope of the 4-DOF parallel observed by
39
rotating 90 in clockwise direction against z-axis from

the origin
Fig. 3.6
39
by rotating 180 in clockwise direction against z-axis

from the origin
Fig. 3.7
40
by rotating 270 in clockwise direction against z-axis

from the origin
Fig. 3.8
Top view of the workspace contour of the 4-DOF
40
parallel manipulator
Fig. 3.9
Bottom view of the workspace contour of the 4-DOF
41
Fig. 4.1
Kinematics diagram of the planar mechanism with
44
actuator pairs (d1, d3)

Fig. 4.2
Kinematics diagram of the vertical mechanism
47
Fig. 4.3
Surface of the objective function f1 (l,k)
55
Fig. 4.4
Surface of the objective function f2 (l,k)
56
Fig. 4.5
Surface of the objective function f(l,k)
56
Fig. 4.6
Values of EM
2
F
at the individual test points in the
57
workspace W in the case when f1 (l,k) is optimized

Fig. 4.7
Values of EM
2
F
at the individual test points in the
58
workspace W in the case when f2 (l,k) is optimized

Fig. 5.1
Free-body diagram of the planar mechanism on the XY-
65
plane
Fig. 5.2
Free-body diagram of the vertical T-mechanism
67
Fig. 5.3
Coordinate frame of the vertical T-mechanism
68
Fig. 5.4
Free-body diagram of the vertical T-mechanism with
73
belt mechanism
Fig. 5.5
Block diagram of the dynamic model of the 4-DOF
74
Fig. 5.6
Simulated actuator positions for case (i)(a) ( , ) =
xiv
77
(89.8, 30 )
Fig. 5.7
Simulated actuator positions for case (i)(b) ( , ) =
77
(89.8, 330 )
Fig. 5.8
Simulated actuator positions for case (ii)(a) ( , ) =
78
(125.7, 150 )
Fig. 5.9
Simulated actuator positions for case (ii)(b) ( , ) =
78
(125.7, 210 )
Fig. 5.10 Simulated trajectories of the end-effector
79
Fig. 5.11 Simulation of the response of p and q with force
80
inputs of case (i)(a)

Fig. 6.1
Prototype of the 4-DOF parallel manipulator
85
Fig. 6.2
Linear AC servo motor of the parallel manipulator
86
Fig. 6.3
High precision linear encoder system
87
Fig. 6.4
Design of high precision passive rotational joint
88
Fig. 6.5
Belt mechanism of the parallel manipulator
88
Fig. 6.6
End-effector of the parallel manipulator
89
Fig. 6.7
Block diagram of the motion control platform of the
91
Fig. 6.8
Block diagram of the current loop control of the servo
92
amplifier
Fig. 6.9
Schematic diagram of the analogue PI current controller
Fig. 6.10 Frequency response of the current loop control of the
93
95
servo amplifier
Fig. 6.11 Difference on the simulated and measured actuator
96
positions for case (,) = (89.8, 30)

96

97

xv
97
Fig. 6.15 Simulated and experimental trajectories of the end-
98
effector
Fig. 7.1
Kinematics design of the 2-DOF planar parallel
101
manipulator
Fig. 7.2
Feedback linearization of the manipulator model
104
Fig. 7.3
Block diagram of the PID computed-torque controller
105
Fig. 7.4
Block diagram of the robust controller
107
Fig. 7.5
Block diagram of the robust learning controller
110
Fig. 7.6
Functional diagram of the learning controller
111
Fig. 7.7
Block diagram of the learning controller
111
Fig. 7.8
Experimental setup of the planar parallel manipulator
113
Fig. 7.9
Model matching of the d1 actuator using PEM
115
identification
Fig. 7.10 Model matching of the d3 actuator using PEM
115
identification
Fig. 7.11 Frequency response of the planar parallel manipulator in
116
d1 joint space
Fig. 7.12 Frequency response of the planar parallel manipulator in
117
d3 joint space
Fig. 7.13 Frequency response of the open-loop system of joint d1
119
119
Fig. 7.15 Sensitivity function of the robust controller of d1
120
actuator
120
actuator
Fig. 7.17 Position tracking of the d1 actuator in the joint space
121
Fig. 7.18 Velocity tracking of the d1 actuator in the joint space
121
122
122
Fig. 7.21 Position error of the d1 actuator in the joint space
123
123
Fig. 7.23 Convergence of maximum dynamic error of d1 and d3
124
actuators using Robust Learning Control
xvi
Fig. 7.24 Position tracking of the end-effector in the Cartesian
125
space determined by the forward kinematics model

Fig. 7.25 Position error of the end-effector in X-direction of the
125
Cartesian space
Fig. 7.26 Position error of the end-effector in Y-direction of the
126
Cartesian space
Fig. 7.27 Position response of the end-effector (grey) and the
127
driving current (black) over the settling interval using

Computed-torque Control
(a) X-direction
(b) Y-direction
128

Robust Control
(a) X-direction
(b) Y-direction
128

Robust Learning Control
(a) X-direction
(b) Y-direction
Fig. 8.1
Feedback linearization of the 4-DOF parallel
133
manipulator model
Fig. 8.2
Block diagram of the PID computed-torque controller of
134
the 4-DOF parallel manipulator

Fig. 8.3
Experimental setup of the 4-DOF parallel manipulator
136
Fig. 8.4
138
identification
Fig. 8.5
138
identification
Fig. 8.6

identification
xvii
139
Fig. 8.7
139
identification
Fig. 8.8
Frequency response of the 4-DOF parallel manipulator
140
in d1 joint space
Fig. 8.9
Frequency response of the 4-DOF parallel manipulator
141
in d3 joint space
Fig. 8.10 Frequency response of the 4-DOF parallel manipulator
141
in d5 joint space
Fig. 8.11 Frequency response of the 4-DOF parallel manipulator
142
in d7 joint space
144
145
145
146
146
actuator
147
actuator
147
actuator
148
actuator
Fig. 8.20 Position tracking of the d1 actuator
148
Fig. 8.21 Velocity tracking of the d1 actuator
149
149
150
150
151
151
152
Fig. 8.28 Position error of the d1 actuator
152
153
153
xviii
154
Fig. 8.32 Position tracking of the end-effector in the Cartesian
155
space determined by the forward kinematics model

Fig. 8.33 Time-based command trajectory of the X, Y and Z
155
directions of the end-effector in the Cartesian space

Fig. 8.34 Time-based command trajectory of the direction of the
156
end-effector in the Cartesian space

Fig. 8.35 Position error of the end-effector in X-direction of the
156
Cartesian space
Fig. 8.36 Position error of the end-effector in Y-direction of the
157
Cartesian space
Fig. 8.37 Position error of the end-effector in Z-direction of the
157
Cartesian space
Fig. 8.38 Position error of the end-effector in -direction of the
158
Cartesian space
Fig. 8.39 Convergence of maximum dynamic error of d1, d3, d5
158
and d7 actuators using Robust Learning Control

160

Computed-torque Control
(a) X-direction
(b) Y-direction
(c) Z-direction
(d) -direction
161

Robust Control
(a) X-direction
(b) Y-direction
(c) Z-direction
(d) -direction
xix
162

(a) X-direction
(b) Y-direction
(c) Z-direction
(d) -direction
xx
List of Tables
Table 4.1 Optimization results of the linkages and the objective
59
functions of the manipulator

Table 4.2 Statistical analysis of the initial kinematics positioning
59
accuracy
Table 4.3 Statistical analysis of the kinematics positioning
59
accuracy optimized by f
60
accuracy optimized by f1
60
accuracy optimized by f2
Table 5.1 Kinematics and dynamic properties of the 4-DOF
76
parallel manipulator for simulation

Table 6.1 Kinematics and dynamic properties of the parallel
90
manipulator
Table 6.2 Electrical characteristics of the linear AC servo motor
92
Table 7.1 Parameters of the robust feedback controller designed by
118
Genetic Algorithm
Table 7.2 Comparison of the maximum absolute dynamic error,
126
settling time and steady-state error of the linear actuators

using different controllers
Table 7.3 Comparisons of the settling time and end-point accuracy
129
(as determined by laser displacement measurement

system) of the planar manipulator using different
controllers and the XY motion stage
Table 7.4 Comparison of the moving mass and the actuator power
of the planar manipulator and the XY stage
xxi
130
Table 7.5 Comparison of the mechanism size, travelling range of
130
the actuators and the workspace envelope volume of the

planar manipulator and the XY stage
Table 8.1 Parameters of the robust feedback controller of the 4-
144
DOF parallel manipulator designed by Genetic

Algorithm
Table 8.2 Analysis of the maximum absolute dynamic error,
159
settling time and steady-state error of the linear actuators

using different controllers
Table 8.3 Analysis of the settling time and end-point accuracy of
the 4-DOF parallel manipulator using different
controllers and the individual XY, Z and Motion Stage
xxii
163
Chapter 1
Introduction
1.1 Background
In accurate and fast positioning application such as semiconductor packaging
industry, position accuracy is essential in assuring the product quality whereas fast
and stable operation enables high production rate to be achieved. Traditional XY
positioning stages have played an important role in semiconductor packaging
machines for translational motion. However, such kind of system consists of several
weaknesses including high payload of movable parts, complexity of mechanical
structure, difficulties in assembly and alignment, and inadequate number of degrees
of freedom for packaging process in one unit of system. Hence, it is essential to look
for a mechanism with light payload, low inertia, high stiffness, simple mechanical
design and multiple degrees of freedom. One suitable choice with all the above
properties lies in parallel manipulators.
The definition of a parallel manipulator given in Merlet [1] is a mechanism made up

of an end-effector with multiple degrees of freedom, and a fixed base, linked together
by at least two independent kinematics chains. The number of actuators is equal to
the number of degrees of freedom (DOF) of the manipulator. Parallel manipulators
are categorized as planar, spherical and spatial in accordance with the number of the
workspace dimensions.
Two advantages of parallel manipulators over the traditional serial robots as high
precision positioning mechanism are low inertia and high stiffness in the structure. In
addition, simple mechanical design and multiple degrees of freedom in a compact
system are benefits of parallel manipulator.
Parallel manipulators can be found in many applications, such as flight simulators,

CNC machining centres, pointing devices and walking machines. However, most of
the existing parallel manipulators are large in size while their accuracy and
repeatability are not adequate for semiconductor packaging applications requiring
micrometer resolution. Hence, there is a need to develop high precision parallel
manipulators with novel kinematics design for the positioning mechanism of
semiconductor packaging system.
1.2 Literature survey

Parallel manipulators have been used in various kinds of industries, after being
proposed by Gough [2] and Stewart [3]. The Stewart-Gough platform (see Fig. 2.1)
is a spatial mechanism in which the moving platform is connected to a fixed base by
six extensible limbs through spherical joints. Each limb is made up of two links that
are connected by a prismatic joint. The Stewart-Gough platform had been modified
and improved for use in multi-axis CNC machining centre [4] and camera-pointing
device [5]. A class of 4-DOF parallel manipulator known as DELTA robot has been
proposed by Clavel [6] as shown in Fig. 2.2, which is mainly used in pick-and-place
machine. Another class of 4-DOF parallel robots is developed by Pierrot [7], [8] as
shown in Fig. 2.3 for high speed handling and machining applications. A third class
of 4-DOF parallel manipulator known as Manta (see Fig. 2.4) and Kanuk are
developed by Rolland [9] for industrial handling purposes. Since all these robots
consist of universal and ball-and-socket joints with multiple DOF, the position
accuracy is hard to assured and it is not suitable for semiconductor packaging
applications.
Good kinematics design is essential for achieving light manipulator structure, simple
kinematics layout and compactness in multiple-DOF manipulator. The design
2
methodology proposed in [10] is mainly used for 6-DOF parallel robot by utilizing
each of the arm elements of the robot with 3-DOF. The design consideration
proposed in [11] is to determine the number of degree of freedom of each link of the
robot to achieve the functional requirements. McCarthy [12] has introduced various
kinds of mechanical linkages and considered the synthesis of these linkages into a
mechanical manipulator. These design approaches provide a sequential design
procedure for building the manipulator. A design approach has been developed in [13]
using the evaluation criteria of static analysis to determine the optimal manipulator
structure. A task-specification design approach has been proposed in [14] using
Stewart platform as the basic module and modified in [15] to increase the number of
different basic modules. A mechatronics design approach is proposed in [16] by
optimizing the manipulator structure with the electrical characteristics of the motor
drive system. However, such approaches are not suitable for our application due to
the lack of design flexibility and concern of industrial constraints. A design approach
is developed in this thesis to fulfill the motion requirements of the parallel
manipulator in size, compactness, joint accuracy, and flexibility of the functional
module assembly.
The size of manipulator workspace and the workspace singularity conditions are two
major criteria for appraising the kinematics design of parallel manipulators. The
geometrical approach, analytical approach and graphical approach are three major
categories of methods for the determination of the manipulator workspace. A
geometrical method has been proposed in [17] to determine the workspace of the
Stewart platform using the boundary intersection of the concentric spheres at
different sections and modified in [18] to determine the workspace of the planar
manipulator. Another geometrical approach has been proposed in [19] to determine
the workspace using the position and mechanism constraints of the manipulator. The
vertex space method has been proposed in [20] to analyze the workspace of the
HexaSlide manipulator. An analytical approach is developed in [21] to determine the
workspace of the 6-DOF parallel manipulator using the analytical solutions of the
boundaries of the three-dimensional workspace. A two-stage minimization approach
has been proposed in [22] to determine the manipulator workspace boundary using
the extreme reachable position of the decoupled open kinematics chains. The Euler-
Rodrigues approach is developed in [23] to determine the workspace of the spherical

and platform mechanisms using the implicit functions of the manipulator singularity
loci. However, complicated geometrical equations have to be solved to determine the
workspace envelope using the approaches in [18-20]. The analytical approaches in
[21-23] require the closed-form kinematics equations of the manipulator to determine
the workspace boundary. Several graphical approaches have been developed to
determine the workspace plots of parallel manipulators using discrete-points
generated by the kinematics equations [24], computer aided design (CAD) tool on
calculating the physical model of the workspace [25] and the boundary search
method [26].
Kinematics optimization is essential for achieving high positioning accuracy and

compactness of a multiple-DOF parallel manipulator. The operational workspace and
structural stiffness have been used as the performance index for most of the existing
optimization algorithms. The operational workspace of the parallel manipulator is
measured by the volume of space that can be reached by the end-effector with all
orientations. The structural stiffness is expressed in terms of the resonant frequency
of the manipulator which may be excited by an input command signal. A global
performance index has been proposed by Gosselin and Angeles [27] to minimize the
kinematics error due to the transformation between joint and Cartesian space. The
global performance index has been modified by Tsai and Joshi [28] and LeguayDurand et al. [29] to optimize the stiffness and the dexterity of the manipulator
workspace. The global dexterity index was recently proposed by Gallant and
Boudreau [30] to optimize the workspace of planar parallel manipulators. The quasiNewton optimization algorithm was proposed by Carretero et al. [31] to minimize
the parasitic motion of a 3-dof spatial manipulator. A nonlinear programming
method had been proposed by Bhattacharya et al. [32] to maximize the rigidity of the
Stewart platform over the desired workspace. A method for optimal kinematics
design of the HexaSlide type parallel manipulator was proposed by Ryu and Cha [33]
using the product of the singular values of the error model for constrained
optimization. The weakness of this approach is that the minimization of the singular
values product does not ensure that the kinematics error is reduced in all directions of
the workspace.
Dynamic modelling of parallel mechanisms is in general difficult and tedious, but is

essential for the motion analysis of parallel manipulators under a given force profile
and for estimating the motion capability in terms of profile tracking and the required
driving forces. Miller and Clavel [34] proposed to use the Lagrange formulation to
determine the dynamic model of the 4-DOF Delta robot. The Euler-Lagrangian
approach has been used by Lee and Shah [35] to determine the dynamic model of a
3-DOF parallel manipulator with translational motion. The Hamilton Principle has
been used by Miller [36], [37] to solve the dynamics of the Delta robot and the
NUWAR robot. The Principle of Virtual Work has been employed by Tsai [38] to
determine the inverse dynamics of a parallel manipulator. Wang and Gosselin [39]
have modified the Principle of Virtual Work to determine the dynamics of a 6-DOF
parallel robot. Dasgupta and Mruthyunjaya [40] modified the Newton-Euler
approach to determine the dynamic equations of the general 6-DOF Stewart platform.
Such an approach is further extended by Dasgupta and Choudhury [41] to other types
of planar and spherical manipulators. The Newton-Euler approach has also been used
by Lee et al. [42] to obtain the dynamic model of a 5-DOF parallel manipulator.
Many different kinds of control techniques have been used in the control of parallel
manipulator. The classical computed-torque control algorithm has been implemented
in [43] and [44] to control the Stewart platform and the DELTA parallel robot. The
main advantages of computed-torque control algorithm are its simple structure and
ease of implementation. However, the computed-torque control method is not
adequate for high precision motion control due to the controller gain limitation.
Various control algorithms have also been used on different types of parallel robots,
such as adaptive PD-feedforward controller [45] for a 6-DOF CNC machining centre,
sliding mode control [46], [47] for a Stewart platform.
Robust control is essential to maintain closed-loop stability and the position accuracy
of the manipulator. Experimental identification method is employed to determine the
dynamic model of the parallel manipulator for the design of robust control system.
Existing approaches to robot model identification include the measurement of motor
input-output data and external force on the motion trajectory [48], application of
Hamilton principle [49] and the maximum-likelihood method [50]. The limitation of
these methods is that the frequency response and the gain variation of the
manipulator cannot be determined. A survey of various robust control methods has
been performed in [51], such as acceleration feedback control [52], computed-torque
plus H control [53], [54], nonlinear approach [55], -synthesis technique [56], fuzzy
logic control [57] and variable structure control [58].
Iterative Learning Control (ILC) is well-suited for enhancing the motion

performance of the parallel manipulator under repetitive motion. Some approaches
are developed including D-type control [59], hybrid control [60,61], previewlearning method [62], high gain feedback design [63], self-tuning control method
[64], Fourier-based learning controller [65-67] and discrete-time iterative learning
controller with partial knowledge on the manipulator joint position and velocity [6871]. However, the methods in [59-64] required an accurate dynamic model and the
approach in [65-67] is sensitive to disturbance. The methods in [68-71] required the
knowledge of the inertia matrix of the manipulator to determine the compensating
torque in the feedforward path.
1.3 Research Objective and Scope

In high-speed and high-precision positioning mechanism, end-point accuracy and
repeatability with fast and stable motion are essential considerations. Furthermore,
the size of the mechanism is required to be compact with respect to its workspace. In
many manufacturing applications (e.g. semiconductor packaging), the positioning
mechanism is usually required to provide 4-DOF motion, where 3 of these motions
are translations in the XYZ space, and the remaining DOF is a rotational motion
against the z-axis.
A robot manipulation and control system consists of the components shown in Fig.
1.1.
Desired
pose
position
Inverse
Kinematics
Desired
actuator
positions
Trajectory
planning and
generation
Estimated
actual pose
position
Control
Algorithm
Control signal
(e.g. current)
Actuator and
robot
mechanism
Actuator
positions
Forward
Kinematics
Fig. 1.1 Block diagram of a typical robot manipulation and control system
Corresponding to the block diagram of Fig. 1.1, a complete design of the system
requires
design
and
analysis
work
in
four
main
areas,
including the
electromechanical system of the robot body, the kinematics and dynamic models, the
robot control algorithm and the motion trajectory generation.
The objective of this research is to develop novel parallel manipulator systems which
provide superior alternative to the traditional XY-stages including significant
improvements in the positioning accuracy and settling time, light moving mass of the
mechanism and reduction in the motor driving power. Compared with a state-of-theart XY-stage used for semi-conductor packaging, our aim is that the desired endpoint positioning error be reduced by an order of magnitude to about one micrometer.
The settling time is desired to be reduced from around 30ms to less than 10ms at the
end of motion. A further motivation for a new design is to evenly distribute the load
due to the total moving mass, including the end-effector and the linkages of the
parallel manipulator to the linear actuators such that the required maximum driving
current of each actuator can be significantly reduced. Our target application of these
parallel manipulator systems is the positioning mechanisms of semiconductor
packaging machines.
In order to achieve the above goal, it is essential to develop the approaches on

manipulator kinematics design, dynamic modelling and controller design to fulfill the
required improvements on motion performance. Most existing designs of parallel
manipulators suffer from size utilization, internal singularity, and complicated
kinematics solution. Kinematics design approach is developed to ensure the parallel
manipulator can perform the required degree-of-freedom motion with compact
7
manipulator size, light manipulator structure, efficient kinematics solution and free
of internal singularity. The workspaces of parallel manipulators have complex shapes
and often cannot be readily determined. A workspace determination approach needs
to be developed to determine the reachable workspace of the end-effector of the
parallel manipulator for visualization and evaluation of the operational volume. For
practical applications, we need to consider manufacturing variations of the
mechanical linkages due to machining capability limitation as these variations can be
magnified at the end-effector position. A kinematics optimization approach will be
developed to minimize the end-effector kinematics error and suppress the kinematics
error magnification effect.
A dynamic model of the parallel manipulator is essential for the purposes of

simulation, analysis (e.g. to determine the dynamical characteristics and the required
driving force of the actuators), and development of model-based control algorithms.
Exploiting the special modular construction of the proposed parallel manipulator
from triangular mechanisms, a specific modeling approach will be developed in this
thesis to obtain a complete mathematical model of the parallel manipulator using
Recursive Newton-Euler method.
Feedback control plays a crucial role in achieving the required end-point position
accuracy. A practical model-based approach is employed to design the PID
computed-torque controller to control the position of parallel manipulators. However,
the computed-torque controller is inadequate to provide the required motion
performance because the controller gain has to be limited to avoid exciting the high
frequency modes of the mechanism. To overcome the weakness of the computedtorque controller, a robust control design method is employed to design a stable, fast
tracking response feedback controller which is less sensitive to high frequency
disturbance. In a repetitive motion between two desired positions, reducing the
dynamic tracking error of the translational joints can help decrease the settling time
of the end-effector at the end of motion. A robust learning controller is developed to
enhance the dynamic tracking performance of the linear actuators of the manipulator
to minimize the settling time of the end-effector and to achieve the desired
positioning accuracy.
Prototypes of the planar parallel manipulator and the 4-DOF parallel manipulator are
developed to assess the practicality of proposed kinematics design and control
strategies. Using the prototypes, the motion performance of the parallel manipulators
using the proposed control schemes are evaluated against traditional XY stages.
Independent
end-point
position
measurements
are
performed
using
laser
displacement measurement system to show that the planar parallel manipulator and
the 4-DOF parallel manipulator outperform existing XY-stages by a significant
margin.
1.4 Contribution
The contributions of this thesis are summarized as follow:
A novel and original design of a 4-DOF parallel manipulator is proposed for

high precision semiconductor packaging application;
As part of the development of the 4-DOF parallel manipulator, a 2-DOF

planar manipulator is developed for planar translational motion application;
Derivation of the kinematics model, Jacobian singularity analysis and the

geometrical method for workspace analysis of the proposed 4-DOF parallel
manipulator;
Development of a novel optimization approach to determine the kinematics

parameters so as to minimize the worst-case kinematics error magnification
of the 4-DOF parallel manipulator within the reachable workspace;
Development of dynamic models for the planar and 4-DOF parallel

manipulators;
Design of computed-torque controllers for the planar and 4-DOF parallel

manipulators;
Development of the experimental identification method for determining the

joint-based dynamic model of the parallel manipulators for the design of the
robust controller;
Development of the robust controller using frequency domain design

approach with parameters optimization using genetic algorithm, and;
Enhancement of the robust controller to the robust learning controller by

developing the learning feedforward control loop using Fast Fourier
Transform (FFT) method for repetitive point-to-point motion.
1.5 Thesis Organization

In this chapter, the background of the parallel manipulator and its application to
various kinds of industry is presented. The literature survey on the parallel
manipulator research, including kinematics design, workspace analysis, dynamic
modelling and control system design are reviewed. The motivation, objectives and
goals of this research are identified for a clear direction on this thesis. The
contributions of this thesis are clearly described to demonstrate the research effort
and output.
In Chapter 2, the design criteria on the motion requirements, positioning accuracy

and manufacturing requirements of the parallel manipulator are presented. The
modular design approach of the high precision 4-DOF parallel manipulator using a
triangular planar mechanism as a building block is developed. The forward
kinematics and inverse kinematics model of the planar mechanism and the 4-DOF
parallel manipulator are derived in closed-form.
In Chapter 3, the Jacobian matrix of the 4-DOF parallel manipulator is derived

relating the velocities of the end-effector and the translational joints. The Jacobian
decoupling approach is then used for the singularity analysis of the parallel
manipulator in both internal and boundary conditions. Such analysis is to ensure that
the proposed manipulator is internal singularity-free within the manipulator
workspace and the workspace boundary is reachable by the end-effector. A novel and
efficient numerical simulation approach using Monte-Carlo method is introduced to
determine the workspace volume of the 4-DOF parallel manipulator.
10
A kinematics optimization approach for high precision positioning accuracy of

manipulators is presented in Chapter 4. Kinematics optimization methods based on
two new objective functions are developed for the 4-DOF parallel manipulator. The
first objective function is designed for minimizing the average of the end-effector
kinematics error over the desired workspace and the second objective function is
developed for suppressing the maximum kinematics error in the chosen operational
workspace. The proposed objective functions yield better kinematics parameters
(with respect to the suppression of kinematics error magnification) compared with
that obtained using the method by Ryu and Cha [33] based on optimization of the
kinematics error along one direction in the desired workspace.
The dynamic modelling of the parallel manipulators is introduced in Chapter 5. The

mathematical model of the planar manipulator and the 4-DOF parallel manipulator
are derived. Recursive Newton-Euler method is used to study the dynamic
characteristics of the manipulator. Numerical simulation is performed to simulate the
response of the dynamic model.
The development of a prototype of the 4-DOF parallel manipulator is presented in

Chapter 6. The geometrical tolerance, structural stiffness and light moving mass of
the manipulator are achieved using high precision mechanical joints and high quality
aluminum alloy linkages design. A PC-DSP motion control platform is established
for the trajectory generation and real-time closed-loop control of the manipulator.
The manipulators are driven by linear AC servo motors and PWM AC servo
amplifier. The PI current controller of the PWM servo amplifier is designed using
frequency domain approach to ensure the gain linearity and the current tracking
bandwidth. The dynamic model of the 4-DOF parallel manipulator is verified with
the prototype in open-loop using the sinusoidal force profiles.
In Chapter 7, the closed-loop control design of the planar manipulator is presented.

The model-based approach of the PID computed-torque controller is firstly employed
for the position control of the manipulator using the model derived in Chapter 5. A
joint-based linear dynamic model is determined experimentally from the prototype
using frequency domain system identification approach, which is then used as a basis
11
for the robust controller design. A joint-based robust learning controller is further
introduced to improve the settling response by reducing the dynamic tracking error
of the manipulator. The performance of these three controllers in terms of settling
time and steady-state positioning accuracy of the end-effector are compared and
independently verified by a high precision laser displacement measurement system.
In Chapter 8, the joint-based control design method for the planar manipulator
described in Chapter 7 is applied to the 4-DOF parallel manipulator. The
mathematical model of the 4-DOF parallel manipulator derived in Chapter 5 is used
for designing the computed-torque controller. The joint-based robust controller is
designed using the linear dynamic model determined by system identification
techniques. The robust learning controller is developed for the 4-DOF parallel
manipulator. Improvements in settling time and steady-state positioning accuracy of
the manipulator are also verified.
Chapter 9 contains a summary of the work described in the thesis. Potential

improvements of the parallel manipulator system and further research work are also
proposed.
12
Chapter 2
Manipulator Kinematics
2.1 Introduction
In this chapter, the design approach and the kinematics analysis of the proposed
parallel manipulator are presented. The design philosophy, criteria on the accuracy of
mechanical joints and actuators and the design-by-module concept are introduced in
the development stage. The manipulator structure is then designed in order to fulfill
the motion requirements. The forward kinematics and inverse kinematics of the
parallel manipulator are derived to obtain the positioning relationship between the
end-effector and the actuators.
2.2 Parallel Manipulator Development

In many manufacturing operations (such as for semiconductor packaging system),
the positioning mechanism is required to provide only 4-DOF motion, where 3 of
these motions are translations in the XYZ space, and the remaining DOF is a
rotational motion against the z-axis. End-point accuracy and repeatability of the
mechanism are essential in assuring product quality. Furthermore, the size of the
mechanism is required to be compact with respect to its workspace.
Parallel manipulators have been used in various kinds of industries ever since the
Stewart platform shown in Fig. 2.1 has been proposed by Gough and Stewart [2],[3].
13
The Stewart platform is modified to form a 4-DOF parallel manipulator called

DELTA by Clavel [6] with a kinematics structure shown in Fig. 2.2. A 4-DOF
parallel manipulator, namely H4 is proposed by Pierrot [7] as shown in Fig. 2.3 for
high speed handling and machining applications. Another class of 4-DOF parallel
manipulator known as Manta is developed by Rolland [9], shown in Fig. 2.4, for
industrial handling processes. The common weakness of the design of these
manipulators is the use multiple DOF passive joints (e.g. universal joints and balland-socket joints) in the linkages. The accuracy of these joints is hardly assured or
the manufacturing costs are otherwise expensive.
Fig. 2.1 Stewart platform
Fig. 2.2 DELTA parallel manipulator
Fig. 2.3 H4 parallel manipulator
Fig. 2.4 Manta parallel manipulator
Task-oriented design approach is developed in this research to design a parallel

manipulator providing the required motion, namely, 4 DOF with three translational
motions in the XYZ space and a rotational motion against the z-axis. We will use
only single DOF joints in the linkages of the parallel manipulator to eliminate the
positioning inaccuracy caused by the multiple DOF mechanical joints. Linear
14
actuator direct-drive approach is selected to drive the manipulator with minimum

number of moving parts due to the actuator. High resolution position feedback device
is mounted adjacent to the linear actuator to determine the actuator position for
estimating the end-effector position. Several design principles are essential for
achieving light manipulator structure, simple kinematics layout and compactness in
the multiple-DOF manipulator. The working envelope of the manipulator is placed at
the central region of the mechanism body. The actuators are located on the fixed base
in a compact kinematics layout to minimize the mechanism size. The distances
between the linkages and the actuators are minimized to avoid geometric error
amplification.
Three types of planar mechanisms are considered as candidates for the basic modules
of the parallel manipulator. A triangular linkage is designed in this thesis by
modifying the Biglide mechanism [72] to reduce the number of passive linkages. The
triangular linkage shown in Fig. 2.5 consists of two linear actuators travelling along
the same axis. The functional point of the triangular mechanism is located at the
combined revolute joint at the end of the two linkages. The functional point in the
mechanism has a limitation that it is not convenient to use in a practical motion
system. Another type of planar mechanism is the H-linkage [73] where the two linear
actuators are placed in parallel with two linkages connected to the centre of the endeffector. The end-effector motion is constrained by a guide rail and two linear motion
blocks which travel along the same axis as the linear actuators. The drawback of this
design is the additional moving mass of the guide rail and the linear motion blocks
increased the total moving mass of the manipulator structure. A V-linkage planar
mechanism is also designed in this thesis by modifying the arrangement of the linear
actuator location of the triangular linkage. The V-linkage shown in Fig. 2.7 suffers
similar disadvantage as the triangular linkage shown in Fig. 2.5 and the size of the Vlinkage mechanism is larger than the triangular linkage.
15
Fig. 2.5 Triangular linkage
Fig. 2.6 H-linkage
Fig. 2.7 V-linkage
The triangular mechanism in Fig. 2.5 is modified by installing an end-effector at the

triangular linkage shown in Fig. 2.8. However, without additional constraint on the
end-effector, a yaw motion will occur. To overcome this problem, a parallelogram
mechanism is installed between the revolute joint of one of the linear actuator and
the revolute joint at the end-effector. The yaw motion of the end-effector is
eliminated by the parallelogram mechanism. The planar mechanism shown in Fig.
2.9 will be used as the basic module for building the 4-DOF parallel manipulator.
Fig. 2.8 Triangular mechanism with end-
Fig. 2.9 Triangular mechanism with end-
effector
effector and parallelogram compensation

mechanism
The configuration of the 4-DOF parallel manipulator is designed to minimize the size
of the manipulator and optimize the compactness to the workspace. The orthogonal
configuration and the parallel configuration of the 4-DOF parallel manipulator are as
shown in Fig. 2.10 and Fig. 2.11 respectively. In these two configurations, the size of
the manipulator in the orthogonal configuration is smaller than the parallel
configuration by half in terms of base area. Another advantage of the orthogonal
configuration is its compact layout reduces the travel distance of the linear actuator
16
required to drive the end-effector to the desired end-point. With the benefits on the
manipulator size and the compactness of the layout, the orthogonal configuration
shown in Fig.2.10 is chosen for the final design of the 4-DOF parallel manipulator.
Fig. 2.10 Two triangular mechanisms

arranged in orthogonal configuration
Fig. 2.11 Two triangular mechanisms

arranged in parallel configuration
In Fig. 2.10, the 4-DOF parallel manipulator is made up of multiple closed triangular
mechanisms. Simple kinematics design is the advantage of the triangular mechanism.
As a result, the kinematics relationships of the mechanism can be readily determined.
As the triangular mechanism consists of only revolute and translational joints, the
joint accuracy due to manufacturing tolerance is more controllable. The triangular
mechanism can also serve as a basic module to provide a 2-DOF motion in the XY
plane. It can also be assembled in different multiple DOF mechanism configuration
to fulfill various motion requirements. The kinematics design of the 4-DOF parallel
manipulator using the multiple triangular mechanisms is next described in section 2.3.
2.3 Manipulator Structure

The kinematics design of the 4-DOF parallel manipulator is shown in Fig. 2.12. Let
OXYZ be the rectangular coordinate system for the base frame. The manipulator
consists of two linear actuator pairs located at (d1, d3) along the X-axis and (d5, d7)
along the Y-axis. The linear actuators are coupled to two 2-DOF triangular planar
mechanisms (referred to as a T-mechanism below) which are used to place the points
17
P(x1, y1) and Q(x2, y2) at desired points on the XY-plane. The T-mechanism can also
provide a 2-DOF planar motion by driving the linear actuators in the same and
opposite directions. Each of the small platforms at the outer joint of the Tmechanisms is constrained by a belt mechanism with two pulleys. One of the pulleys
is mounted with the small platform which can rotate relative to the shaft connecting
the revolute joints of l2 and l4. The other pulley is fixed to the shaft of the revolute
joint on d1 and d7 and cannot be rotated. This design is equivalent to the
parallelogram structure (as shown in section 2.2) to constraint the small platform of
the T-mechanism to have only translational motion with the linear actuators but not
rotation on the XY-plane.
Fig. 2.12 Kinematics design of the 4-DOF parallel manipulator
Another T-mechanism is constructed on top of the small platforms at the outer ends
of the two planar T-mechanisms, with two rotational joints at each of the points P
and Q. The third T-mechanism has motion restricted to the vertical plane. A platform
mounted rigidly at the top of the vertical T-mechanism serves as the end-effector.
The surface of the platform is constrained to the horizontal plane by a belt
mechanism (which is also equivalent to the parallelogram structure in the planar T18
mechanism) but can otherwise rotate about the Z-axis as the positions of P and Q are
manipulated.
Hence, 4-DOF motion can be obtained by driving the small platforms at the outer
joint of the T-mechanisms in the same and different directions of x and y axes. The
x-direction motion of the end-effector can be achieved by driving the small platforms
in the same x1 and x2 directions simultaneously, while the y-direction motion is
obtained by synchronizing the motion of the small platforms y1 and y2 in the same
directions. The z-direction motion is achieved by controlling the distance between P
and Q, while rotational motion in is obtained by controlling the orientation of the
line joining P and Q.
2.4 Kinematics of the Planar Parallel Manipulator

In this section, the forward kinematics and inverse kinematics of the planar parallel
manipulator are derived to determine the mathematical relationship between the
positions d1, d3, d5, d7 of the linear actuators and the positions of the points P(x1, y1)
and Q(x2, y2).
Let the positions d1, d3, d5, d7 satisfy the conditions:
0 < d1min d1 d1max < d3min d3 d3max

0 < d7min d7 d7max < d5min d5 d5max
and
(2.4)
In Fig. 2.12, the linear actuators with positions d1 and d3 travel along the X-axis on
the same linear motion guide and the other two actuators pair with positions d5 and d7
travel along the Y-axis on another linear motion guide. The conditions of equation
(2.4) must be fulfilled to avoid the linear actuators crashing into each other on the
linear motion guide during dynamic motion. Two mechanical stoppers, one is
installed at the position of d1max and the other is installed at the position of d7max to
limit the travel of the linear actuator pairs.
19
2.4.1 Forward Kinematics

In the forward kinematics of the planar parallel manipulator, the positions of the
points P(x1, y1) and Q(x2, y2) are determined from the given positions of the linear
actuators d1, d3, d5 and d7. In Fig. 2.12, suppose the lengths of the linkages of the
planar triangular mechanisms are identical, i.e. l2=l4=l6=l8=l, then the coordinates of
the point P(x1, y1) and Q(x2, y2) are related to the linear actuator positions by:
x1 =
1
( d1 + d3 )
2
(2.5)
y1 = l 2
1
2
( d3 d1 ) + d y
4
(2.6)
x2 = l 2
1
2
( d5 d 7 ) + d x
4
(2.7)
y2 =
1
( d5 + d 7 )
2
(2.8)
where dy is the distance between the point P(x1, y1) and the coupling joints of l2 and
l4 and dx is the distance between the point Q(x2, y2) and the coupling joints of l6 and l8.
2.4.2 Inverse Kinematics

The inverse kinematics of the planar parallel manipulator will be needed to
determine the positions of the linear actuator when the coordinates of the point P(x1,
y1) and Q(x2, y2) are given.
From equations (2.5), the position of the linear actuator d3 can be expressed in terms
of d1 and x1 as
d 3 = 2 x1 d1
20
(2.9)
Equation (2.6) is expanded and modified to determine the positions of the linear
actuator d1 and d3 as:
(y d )
1
= l2
1
2
( d3 d1 )
4
(2.10)
Substituting equation (2.9) into (2.10) the position of the linear actuator d1 can be
expressed as
d1 = x1 l 2 ( y1 d y )
The position of the linear actuator d3 can be determined by substituting d1 into

equation (2.9) as
d 3 = x1 l 2 ( y1 d y )
From the kinematics design shown in Fig. 2.12 and equation (2.4), d3 is larger than d1,
and hence the solution for d1 and d3 can be uniquely determined as
d1 = x1 l 2 ( y1 d y )
d 3 = x1 + l 2 ( y1 d y )
(2.11)
(2.12)
A similar approach can be applied to determine the positions of the linear actuators
d5 and d7 along the y-axis, yielding
d5 = y2 + l 2 ( x2 d x )
21
(2.13)
d7 = y2 l 2 ( x2 d x )
(2.14)
2.5 Kinematics of the 4-DOF Parallel Manipulator

The forward and inverse kinematics of the planar parallel manipulator between the
positions of the linear actuators d1, d3, d5, d7 and the coordinates of the small
platforms P(x1, y1) and Q(x2, y2) are given in equations (2.5) to (2.8) and (2.11) to
(2.14). These equations will now be used to determine the forward kinematics and
inverse kinematics of the 4-DOF parallel manipulator between the linear actuators
and the coordinate of the end-effector denoted as G(x,y,z,).
We note that by the kinematics design of the 4-DOF parallel manipulator, the
coordinates P and Q of the small platform of the two planar triangular mechanisms
must fulfill the conditions as
x1 > x2
and
y1 < y2
(2.15)
2.5.1 Forward Kinematics

In the forward kinematics of the 4-DOF parallel manipulator, the position of the endeffector G(x,y,z,) is determined from the coordinates P(x1,y1) and Q(x2,y2). In Fig.
2.12, let the lengths of the vertical linkages of the 4-DOF parallel manipulator be
identical, i.e. k10=k12=k. The end-effector position G(x,y,z,) can be expressed in
terms of the coordinates of x1, y1, x2 and y2 as:
x=
1
( x1 + x2 )
2
(2.16)
y=
1
( y1 + y2 )
2
(2.17)
22
z= k
( x x ) + ( y2 y1 )
2 1
4
+ dz
y2 y1
x2 x1
= tan 1
(2.18)
(2.19)
where x1, y1, x2 and y2 are obtained from the position of the linear actuators using
equations (2.5) to (2.8).
2.5.2 Inverse Kinematics

The inverse kinematics of the 4-DOF parallel manipulator can be readily obtained to
determine the positions of the linear actuator when the coordinates of the endeffector position G(x,y,z,) are given.
Re-arrange equations (2.18) and (2.19) as
( y2 y1 )
2
2
= 4 k 2 ( z d z ) ( x2 x1 )
( y2 y1 )
= tan 2 ( x2 x1 )
(2.20)
(2.21)
The positions of x1 and x2 can be expressed in terms of z and coordinates of the

end-effector as
( x2 x1 )
2
4 k 2 ( z d z )
=
2
(1 + tan )
(2.22)
From equation (2.16), x2 is rearranged and expressed in terms of x1. Substitute into
(2.22) to determine x1 as
23
k 2 ( z d z )2
x1 = x
2
(1 + tan )
Applying equation (2.16), x1 can be expressed in terms of x2 and substitute into

equation (2.22) to determine x2 as
k 2 ( z d z )2
x2 = x
2
(1 + tan )
Noting that
1
1
=
= cos 2 and applying the kinematics condition in
2
2
1 + tan sec
(2.15), x1 and x2 can be uniquely determined as
x1 = x + k 2 ( z d z ) cos
x2 = x k 2 ( z d z ) cos
(2.23)
(2.24)
Similarly, the inverse kinematics for y1 and y2 are given by
y1 = y k 2 ( z d z ) sin
y2 = y + k 2 ( z d z ) sin
(2.25)
(2.26)
The positions of the linear actuators d1, d3, d5 and d7 can be expressed in terms of the
positions of the coordinates P(x1, y1) and Q(x2, y2) using equations (2.11) to (2.14).
This completes the inverse kinematics expressing d1, d3, d5 and d7 in terms of the
end-effector position G(x,y,z,)
24
2.6 Summary
The development and kinematics analysis of a novel 4-DOF parallel manipulator is
presented in this chapter. A modular design approach is taken to design the
manipulator using triangular mechanisms as building blocks to provide the required
motion while ensuring that the kinematics structure has inherent positioning accuracy
without unreasonably demanding requirements on manufacturing of mechanical
joints. The structural simplicity of the proposed parallel manipulator also leads to an
efficient algorithm for kinematics determination. With the benefits of the kinematics
design, multiple solutions of the inverse kinematics of many common parallel
manipulators are eliminated, so that the solution to the inverse kinematics of the
proposed parallel manipulators can be uniquely determined by closed-form equations.
25
Chapter 3
Jacobian, Singularity and Workspace
Analysis
3.1 Introduction
Workspace singularity is one of the limitations in many existing parallel
manipulators. In the reachable workspace of the parallel manipulator, it is desirable
to avoid internal singularity such that every position within the workspace can be
achieved by the end-effector. In this chapter, the manipulator Jacobian is derived to
determine the relationship between the velocities of the linear actuators and the
velocity of the end-effector. The singularity conditions of the parallel manipulator
can then be determined using the Jacobian matrix to ensure that no singularity exists
within the reachable workspace of the manipulator. A numerical simulation method
is developed to determine the workspace envelope of the parallel manipulator. This
method leads to an efficient approach for determining the irregular workspace of the
manipulator, especially when the analytical solution of the manipulator boundary
cannot be obtained.
3.2 Derivation of Jacobian Matrix

The Jacobian matrix of the 4-DOF parallel manipulator is derived to determine the
mapping between the velocities of the linear actuators and the velocity of the end26
effector. By the design of the parallel manipulator, it is driven by four linear

actuators and the end-effector has 4-DOF motion. The positions of the linear
actuators can be represented by a vector q = [ d1
d3
d 7 ] and the end-effector
d5
position can be expressed as another vector X = [ x
z ] . The Jacobian
matrix of the 4-DOF parallel manipulator is defined as
X = Jq
where X = x
q = d1
d3
d5
y
(3.1)
T
z is the velocity vector of the end-effector and
T
d7 is the velocity vector of the linear actuators. In the parallel
manipulator, the number of actuators is equal to the number of degree-of-freedom

without redundancy. The Jacobian matrix can therefore be expressed as a 4x4 square
matrix:
x
d
1
y
d
1
J =
z
d1

d1
x
d3
x
d5
y
d3
y
d5
z
d3
z
d5
d3
d5
x
d 7
y
d 7
d 7

d 7
(3.2)
By (2.16) to (2.19) and expressed the coordinates of x1, y1, x2 and y2 in terms of the
actuator positions using the kinematics equations (2.5) to (2.8), the elements of J are
given by
x
x 1
=
=
d1 d 3 4
(3.3)
x
x d 7 d5
=
=
d5
d 7
4 p1
(3.4)
27
z
=
d1
z
=
d3
z
=
d5
z
=
d 7
y
y d 3 d1
=
=
d1
d3 4 p2
(3.5)
y
y 1
=
=
d5 d 7 4
(3.6)
p1 d1 d 3 +
2 p3 ( d 3 d1 )
p2
+ 2d x
(3.7)
8 4k 2 p4 2 p32
2 p3 ( d1 d 3 )
p1 d1 d 3 +
p2
+ 2d x
(3.8)
8 4k 2 p4 2 p32
p2 d 7 d 5
2 p4 ( d 7 d 5 )
p1
+ 2d y
(3.9)
8 4k 2 p4 2 p32
p2 d 7 d 5
2 p4 ( d 5 d 7 )
p1
8 4k 2 p4 2 p32
p4 d1 d3 p3
+
2 p2
p4
=
p32 + p4 2
d1
p4 d 3 d1 p3
+
2 p2
p4
=
d3
p32 + p4 2
=
d5
=
d 7
p3 ( d 7 d 5 )
p4
1
2
p4 p1
p32 + p4 2
p3 ( d5 d 7 )
p4
1
2
p4 p1
p32 + p4 2
+ 2d y
(3.10)
(3.11)
(3.12)
(3.13)
(3.14)
where in the partial derivative components of the Jacobian matrix given in (3.3) to
(3.14),
28
p1 = 4l 2 ( d 5 d 7 )
p2 = 4l 2 ( d3 d1 )
p3 = d 5 + d 7 4l 2 ( d 3 d1 ) 2d y
p4 =
4l 2 ( d 5 d 7 ) d1 d 3 + 2d x
)2
and
)2
To determine the singularity conditions of the parallel manipulator, one can attempt
to compute the determinant of the Jacobian matrix (3.2). However, the expression of
the determinant is extremely tedious and as a result the solutions for the singularity
conditions are difficult. To overcome the difficulty in determining the singularity
conditions of the parallel manipulator, a Jacobian decoupling approach is employed
to decompose the Jacobian matrix in (3.2) into two 4x4 matrices so that the
determinant of the two matrices can be readily computed separately.
In the Jacobian decoupling approach, a position vector of X p = [ x1
y1
x2
y2 ] is
defined to express the positions of the small platforms of the two T-mechanisms as
shown in Fig. 2.12. The Jacobian matrix J q of the linear actuators and the small
platforms can be expressed as
X p = J q q
x1
d
1
y1
d
1
Jq =
x2
d1
y2
d1
x1
d3
x1
d 5
y1
d3
y1
d 5
x2
d3
x2
d 5
y2
d3
y2
d 5
29
(3.15)
x1
d 7
y1
d 7
x2
d 7
y2
d 7
(3.16)
where X p = [ x1
y1
x2
y 2 ] is the velocity vector of the small platforms of the T-
mechanisms. The velocity relationship between the small platforms and the endeffector is expressed by another Jacobian matrix J H as
X = J H X p
x
x
1
y
x
1
JH =
z
x
1

x1
x
y1
x
x2
y
y1
y
x2
z
y1
z
x2
y1
x2
x
y2
y
y2
z
y2

y2
(3.17)
(3.18)
By (3.15) to (3.18), the Jacobian matrix of the 4-DOF parallel manipulator in (3.1)
can be written as:
X = J H J q q
(3.19)
From (2.5) to (2.8), the positions of (x1, y1) are only related to the positions of the
linear actuators d1 and d3 while the positions of (x2, y2) are only related to the
positions of the linear actuators d5 and d7. Hence, the Jacobian matrix J q in (3.16)
takes the form:
30
x1
d
1
y1
d
1
Jq =
x1
d 3
y1
d 3
x2
d 5
y2
d 5
x2
d 7
y2
d 7
(3.20)
Similarly, from (2.16) to (2.19), the position of the end-effector in x-direction is

related to x1 and x2 and the position in y-direction is related to y1 and y2. The Jacobian
matrix J H in (3.19) is simplified as:
x
x
1
0
JH =
z
x
1

x1
x
x2
y
y1
z
y1
z
x2
y1
x2
y
y2
z
y2

y2
(3.21)
From the Jacobian matrices J q and J H expressed in (3.20) and (3.21), several
elements are equal to zero due to the simplicity of the geometrical structure. The
determinant of the Jacobian matrices of J q and J H can be computed as:
y1
d 3
x
det ( J q ) = 1 0
d1
0
x2
d5
y2
d 5
x2 x1
d 7 d3
y2
d 7
31
y1
d1
0
x2
d 5
y2
d 5
x2
d 7
y2
d 7
(3.22)
y1
x z
det ( J H ) =
x1 y1

y1
0
z
x2
x2
y2
z x
+
y2 x2

y2
x1

x1
y
y1
z
y1
y1
y2
z
y2

y2
(3.23)
3.3 Singularity Analysis

One of the limitations of a parallel manipulator is its singularity condition. From the
definition of manipulator singularities by Sciavicco and Siciliano [74], there are 2
categories of singularities. The first category is internal singularity. Such singularity
occurs within the workspace of the parallel manipulator, which reduces the mobility
of the mechanism. The second category is boundary singularity. Such singularity
occurs when the end-effector is on the surface of the workspace boundary. The
boundary singularity can also occur when the actuators reach the limiting positions.
Since the number of degree-of-freedom of the manipulator and the number of

actuators are equal to F = 4, the Jacobian matrices must be 4x4 with normal rank = 4.
The singularity of the parallel manipulator occurs when
Rank ( J H J q ) < 4 min Rank ( J H ) , Rank ( J q ) < 4
(3.24)
The conditions where the rank of JH and Jq are less than 4 can be determined by
equating (3.22) and (3.23) to zero.
det ( J H ) =
1
2
2 4k 2 ( x1 x2 ) ( y1 y2 )
det ( J q ) =
4
( d1 d3 )( d7 d5 )
2
2
4l 2 ( d1 d3 ) * 4l 2 ( d 7 d5 )
32
(3.25)
(3.26)
In (3.25), the determinant of J H is not be equal to zero in all cases. The determinant
of J q becomes 0 when:
d1 = d 3 or d 5 = d 7
(3.27)
Equation (3.27) defines the conditions when the internal singularity occurs. Since the
linear actuators cannot be driven to the same position simultaneously, condition (3.27)
will never occur. Hence, there is no internal singularity in the actual workspace of
the proposed parallel manipulator.
The boundary singularity of the manipulator can be determined by the geometrical

characteristic equations of the planar and vertical linkages. Consider the following
physical conditions of the manipulator:
y1 = y2 2k and x1 = x2
(3.28)
d 3 = d1 + 2l or d5 = d 7 + 2l
(3.29)
(3.28) and (3.29) indicate the boundary singularity conditions of the 4-DOF parallel
manipulator. This singularity can be avoided in the manipulator design by limiting
the travelling range of each linear actuator.
3.4 Determination of the Workspace Envelope

The reachable workspace of the parallel manipulator is defined as the reachable
region of the end-effector within the volume of the mechanism. Limited workspace is
one of the drawbacks of many existing design of parallel manipulators. The
simplicity of kinematics design helps to reduce the manipulator size and avoid the
singularity of the manipulator workspace. However, the kinematics design results in
an irregular workspace. It is necessary to develop an efficient method to generate the
3D workspace for visualization and workspace determination. In the 4-DOF parallel
manipulator designed in this project, the kinematics equations cannot be solved in
33
closed form to give the analytical solutions for the boundary of the workspace.
Numerical simulation is employed to determine the workspace of the parallel
manipulator. The 3D Delaunay Tessellation method is evaluated to determine the
manipulator workspace. However, the output of the 3D Delaunay Tessellation
method is difficult to manipulate and the result is not consistent in different
configurations. To overcome the limitation of 3D Delaunay Tessellation method, a
modified Monte-Carlo simulation and edge detection approach is developed to
determine the workspace envelope of the 4-DOF parallel manipulator.
3.4.1 3D Delaunay Tessellation Method

The workspace boundary of the two planar mechanisms illustrated in Fig. 2.12 can
be determined using the kinematics equations (2.5) to (2.8). The boundary coordinate
pairs of the planar mechanisms are used for generating the data points of the 3D
workspace of the manipulator. The Delaunay tessellation creates a set of lines
connecting each data point on the 3D space to its natural neighbours to form a set of
triangles. In these triangle sets, no data points are contained in any triangle's
circumscribed circle. The triangle sets of all the data points are assembled to form the
workspace of the manipulator. The VRML model of the manipulator workspace is
generated for analyzing the reachable volume of the end-effector.
The desired workspace is specified and the traveling range of the linear actuators is
defined in the design stage. The dimensional parameters are then determined by trial
and error based on the kinematics equations to ensure that the desired workspace can
be achieved by the specified traveling range of the linear actuators. The set of
dimensional parameters is used to generate the workspace boundary of the 4-DOF
parallel manipulator. Fig. 3.1 shows the workspace envelope of the parallel
manipulator determined by the 3D-Delaunay tessellation method and observed from
the origin of the parallel manipulator.
With the 3D Delaunay Tessellation method, the workspace of the parallel

manipulator can be generated regardless of the lack of closed-form analytical
solutions. The pre-processing operations of the workspace data points, such as
34
coordinate arrangement, data segmentation can be minimized. However, the form of

the workspace envelope generated using this method is difficult to interpret due to
the limited visibility and the shape of the workspace envelope is not clear. The
workspace envelope generated by the Delaunay Tessellation method is not a closed
and complete volume. Hence, it is necessary to develop a numerical workspace
determination method which can overcome the weakness of the 3D Delaunay
Tessellation method. The Monte-Carlo simulation method with edge detection
algorithm for workspace boundary extraction is developed to determine the
workspace envelope of the 4-DOF parallel manipulator.
Fig. 3.1 Workspace envelope of the 4-DOF parallel manipulator determined by 3D

Delaunay Tessellation method
3.4.2 Monte-Carlo Simulation Method

In the Monte-Carlo simulation method, an initial set of data points within the
workspace of the parallel manipulator (in this case is over 1,000,000 points) is
generated by incrementally adjusting the position of the linear actuators and
computing the end-effector position using the kinematics equations (2.5) to (2.8) and
35
(2.16) to (2.19). The workspace volume of the parallel manipulator is sliced into
multiple planar surfaces parallel to the xy-plane. The vertical distance between the
intervals of each planar surface is identical. To maintain the precision of the
workspace envelope of the parallel manipulator, the number of planar surface h
sliced from the vertical axis of the workspace is determined by
max ( zi ) min ( zi )
h = round
Zd
(3.30)
where zi is the position of the data point i in z-axis and Z d is the desired vertical
distance between each planar surface. In this simulation, the value of Z d is defined as
0.4mm. The related data points of the workspace envelope of the parallel
manipulator located in the specific planar surface are acquired. These data points are
grouped into a subset Ph = ( xh ( i ) , yh ( i ) ) where xh ( i ) and yh ( i ) are the coordinates
of the ith data point of the workspace within the hth planar surface. A grid table is
generated to cover all the positions of the data points at each specific planar surface
and shown in Fig. 3.2.
Fig. 3.2 Grid table of mapping the x-y coordinates of the data points in the
manipulator workspace at the specific planar surface
36
The x-y coordinates of each data point is mapped into the corresponding grid in the
grid table determined by
gc ( h ) =
gw ( h ) =
xh (i ) min ( xh ( i ) )
rx ( h )
max ( yh ( i ) ) yh ( i )
ry ( h )
(3.31)
(3.32)
where gc ( h ) is the number of column and g w ( h ) the number of row of the grid table
at the planar surface h which represented the mapping of the ith data point, rx ( h ) and
ry ( h ) are the column resolution and the row resolution of the grid table (in unit of
mm). From the results of the singularity analysis shown in section 3.3, there is no
internal singularity within the workspace of the manipulator. The edge detection
algorithm is used to determine if any data point located in the specific grid. If there
has data point located in the specific tested grid, the grid content will be set to one,
otherwise the grid content will remain zero. Since the content of the grid table is
binary, it is equivalent to a black-and-white image. Fig. 3.3 shows the mapping of the
grid table with contents to a binary image.
Fig. 3.3 Mapping of the grid table of the workspace data points to a binary image
From the binary image of the grid table, the boundary of the workspace of the
specific planar surface is determined by the zero-cross edge finding algorithm [75]. It
is necessary to arrange the boundary grid points of the workspace in the specific
37
planar surface so that it can be detected by the edge finding algorithm. The Canny
Path Finding Method is employed to determine the boundary of the binary image
which represents a cross section of the workspace envelope.
The workspace envelope in each of the planar surface sliced from the z-axis is
required to connect in order to create a closed workspace volume using volumetric
surfing method. In the volumetric surfing method, the number of data points in the
workspace envelope of each planar surface is required to be identical. Linear
interpolation method is used to generate the required number of data points to form
the closed workspace boundary in the specific planar surface.
The workspace envelope of the parallel manipulator can be determined by the

volumetric surfing method using the interpolated data points of the workspace
boundary in the specific planar surface with the corresponding vertical position in zaxis. Figures 3.4-3.7 show the workspace envelope of the 4-DOF parallel
manipulator which determined by the Monte-Carol simulation method and observed
from different orientations. The workspace contours of the parallel manipulator
observed from the top view and the bottom view are showin in Fig. 3.8 and Fig. 3.9
respectively.
Fig. 3.4 Workspace envelope of the 4-DOF parallel manipulator observed from the
origin
38
Fig. 3.5 Workspace envelope of the 4-DOF parallel observed by rotating 90 in

clockwise direction against z-axis from the origin
Fig. 3.6 Workspace envelope of the 4-DOF parallel manipulator by rotating 180 in
39
Fig. 3.7 Workspace envelope of the 4-DOF parallel manipulator by rotating 270 in
Fig. 3.8 Top view of the workspace contour of the 4-DOF parallel manipulator
40
Fig. 3.9 Bottom view of the workspace contour of the 4-DOF parallel manipulator
From the workspace envelope of the 4-DOF parallel manipulator shown in Figures
3.4-3.7 and the workspace contour shown in Figures 3.8-3.9, the Monte-Carlo
simulation with edge detection method developed in this research can determine a
closed and completed volumetric workspace of the parallel manipulator as compared
to the results of the 3D Delaunay Tessellation method shown in Fig. 3.1. The
visibility of the workspace envelope is improved significantly and the workspace
envelopes at different vertical positions are indicated in the top view and the bottom
view of the contour plots. The coordinates of the workspace enveloped determined
from the Monte-Carlo simulation method can be used for further analysis on the
kinematics design and workspace modification.
3.5 Summary
The Jacobian matrix of the 4-DOF parallel manipulator is derived for singularity
analysis. With the benefits of the Jacobian decoupling approach, the Jacobian matrix
is decomposed into two 4x4 matrices which simplified the determination of the
singularity conditions of the parallel manipulator. The results of the singularity
analysis demonstrated that there is no internal singularity within the workspace of the
parallel manipulator. This shows that the design criterion on reachable workspace is
41
fulfilled and each position within this workspace can be reached by the end-effector.
The Monte-Carlo simulation method is developed to determine the workspace
envelope of the 4-DOF parallel manipulator. Significant improvement is obtained on
the performance of the workspace determination compared to the 3D Delaunay
Tessellation method. Using the Monte-Carlo simulation method, closed-form
analytical solution for the workspace boundary is not required for determining the
workspace envelope of the manipulator.
42
Chapter 4
Kinematics Optimization
4.1 Introduction
In high precision positioning mechanisms, end-effector accuracy and repeatability
are essential in assuring the product quality. Furthermore, the size of the mechanism
is required to be compact with respect to the workspace. Under these constraints, the
structural design of the manipulator may cause the manufacturing variations of the
mechanical linkages to be magnified at the position of the end-effector. The accuracy
of the linkages is constrained by the manufacturing capability of the machines. To
overcome the problem of error magnification, a kinematics optimization approach for
high precision positioning accuracy of parallel manipulators is developed in this
chapter to determine the kinematics parameters so as to minimize the kinematics
error within a prescribed reachable workspace volume. The error magnification
matrix of the 4-DOF parallel mechanism is derived and two objective functions for
minimization of the end-effector kinematics error are developed. The results of the
optimization are given together with the statistical analysis of the kinematics error
distribution of the end-effector in an operational workspace.
4.2 Derivation of Error Manipulation Matrix

The end-effector kinematics error of a manipulator is caused by variations in
machining accuracy of linkages. In most manipulators, the structural design results in
43
an error magnification from the linkage variation to the end-effector position.

However, the machining accuracy of the mechanical linkages is constrained by the
capability of the manufacturing machines. An Error Manipulation Matrix is derived
that transforms the accuracy variations of the mechanical linkages to the end-effector
kinematics error. The Error Manipulation Matrix is then used to optimize the length
of the mechanical linkages to minimize the kinematics error at the end-effector
within the desired reachable workspace.
First, the Error Manipulation Matrix of the two planar mechanisms is derived. A
kinematics diagram of the planar mechanism with the actuator pair (d1, d3) is shown
in Fig. 4.1.
Fig. 4.1 Kinematics diagram of the planar mechanism with actuator pairs (d1, d3)
In Fig. 4.1, assume the length of the mechanical linkages of the planar mechanism
are identical, i.e. l=l2=l4. Suppose the length of the linkage l2 is changed to (l + l2)
where l2 represents the manufacturing variation of l2, and l4=l, the position of x1
which is the new position of x1 caused by the variation of l2 is expressed as
( x1 d1 ) ( d3 x1 )
= ( l + l2 ) l 2
Simplifying equation (4.1), x1 can be expressed as
44
(4.1)
x1 =
( d1 + d3 ) +
2
l l2
( d3 d1 )
(4.2)
The position of x1 can also be expressed as x1 = x1 + x1 . Substitute into equation

(4.2), the variation of the position x1 can be determined
x1 =
l l2
( d3 d1 )
(4.3)
where x1 is the position difference of x1 due to the variation of the linkage l2. In the
y-direction, the position of y1 which is the new position of y1 caused by the variation
of l2 is expressed as
y12 = l 2 ( d3 x1 )
(4.4)
Substitute equation (4.2) into (4.4) and simplified the expression yields
y1 = l
(d d )
3 1
4
+ l l2 = ( y1 + y1 )
(4.5)
As the value of y1 is in the unit of micron, the second order variation of y1 can be
neglected and y12 is expressed as y12 y12 + 2 y1 y1 . Substitute this expression into
equation (4.5) and using equation (2.5) with y equals to zero
l l2
y1 =
2
(d d )
l2 3 1
(4.6)
The position variation (x1, y1) caused by the variation of the length of the linkage l2
is now determined. In the case where the length of the linkage l4 changes to (l + l4)
where l4 is the manufacturing variation of l4 and l2=l, it can be shown by an
argument similar to the above that
45
x1 =
l l4
( d3 d1 )
(4.7)
l l4
y1 =
2
(d d )
l2 3 1
(4.8)
Applying similar arguments to determine the error kinematics of the planar

mechanism with the actuator pair (d5, d7) caused by the length variation of the
mechanical linkages l6 and l8, the error manipulation matrix of the two planar
mechanisms can be expressed as
l
( d3 d1 )
2
x1
( d3 d1 )
2
y 2 l
4
1 =
x2
0

y2
l
( d3 d1 )
l
(d d )
l2 3 1
0
2 l2
( d5 d 7 )
4
l
( d5 d7 )
l
2
2 l2
( d5 d 7 )
4
l
( d5 d 7 )
l2
l (4.9)
4
l6
l
8
where l6 and l8 are the length variation of the mechanical linkages l6 and l8 of the
planar mechanism with the actuator pair (d5, d7) on Y-axis respectively. The error
kinematics matrix in (4.9) will be used to determine the error kinematics of the endeffector in terms of the position variation of the small platform of the two planar
mechanisms and the length variation of the vertical linkages.
Fig. 4.2 shows the simplified kinematics diagram of the vertical mechanism. The
end-effector position ( x, y, z , ) are defined as
x = x + x , y = y + y , z = z + z and = +
46
where (x, y, z, ) are the original coordinates of the end-effector position and (x, y,
z, ) are the kinematics error of the end-effector due to the variation of the small
platform position and the length of the vertical linkages.
Fig. 4.2 Kinematics diagram of the vertical mechanism
Assume the variation of the length of the vertical linkages is zero, i.e. k10= k12=0,
the position variation of the end-effector in x and y directions is effected by the
position variation of (x1, y1) and (x2, y2) as
x + x =
y + y =
( x1 + x1 + x2 )
2
( y1 + y1 + y2 )
2
and
x + x =
and
y + y =
( x1 + x2 + x2 )
2
( y1 + y2 + y2 )
2
Using equation (2.16) and (2.17), we can solve x and y as
x=
x1
2
and
47
x=
x2
2
(4.10)
y=
y1
2
and
y=
y2
(4.11)
For the position variation in z direction, using equation (2.18) and neglecting the
higher order variation implies that z 2 z 2 + 2 z z . The kinematics error of z-axis
caused by the position variation of (x1, y1) and (x2, y2) can then be expressed as
x22 2 x2 ( x1 + x1 ) + ( x1 + x1 )2 + ( y2 y1 )2
z 2 + 2 z z = k 2
,
4
z 2 + 2 z z = k 2
( x2 x1 )
2
+ y22 2 y2 ( y1 + y1 ) + ( y1 + y1 )
,
4
( x2 + x2 )2 2 x1 ( x2 + x2 ) + x12 + ( y2 y1 )2
z 2 + 2 z z = k 2
and
4
z 2 + 2 z z = k 2
( x2 x1 )
2
+ ( y2 + y2 ) 2 y1 ( y2 + y2 ) + y12
.
4
Solving z in terms of x1, y1, x2, y2 and neglecting the higher order position
variation yields
z =
( x1 x2 ) x ,
z =
( x2 x1 ) x
4z
4z
z =
and
( y1 y2 ) y ,
z =
4z
( y2 y1 ) y
4z
(4.12)
In the case when the vertical linkages k10 and k12 have variations as shown in Fig.
4.2(b), the kinematics error of the end-effector in the z-direction can be determined
as
48
z =
k k10
2
( y y ) + ( x2 x1 )
2 k2 2 1
4
and z =
k k12
2
( y y ) + ( x2 x1 )
2 k2 2 1
4
(4.13)
The kinematics error of the end-effector in the x and y directions due to the length
variation of k10 and k12 are expressed as
x=
x=
k k10
2
( y2 y1 ) + ( x2 x1 )
k k12
2
( y2 y1 ) + ( x2 x1 )
cos
and
cos
and
y=
y=
k k10
2
( y2 y1 ) + ( x2 x1 )
k k12
2
( y2 y1 ) + ( x2 x1 )
sin (4.14)
sin (4.15)
The kinematics error in the direction is determined by differentiating equation

(2.19) with respect to (x1, y1) and (x2, y2) such that
( y2 y1 )
( x2 x1 )
( y1 y2 )
=
x1 x2 )
2
2
2 (
( x2 x1 )
( x2 x1 ) + ( y2 y1 ) ( x2 x1 )
2
(4.16)
From equations (4.10) and (4.16), the transformation of the kinematics error of the
T
end-effector E = [ x y z ] from the position variation of the small

platforms of the planar mechanisms (x1, y1) and (x2, y2) and the length variation
of the vertical linkages due to the manufacturing error (k10, k12) can be obtained.
Using the matrix in (4.9) to express the position variation of the small platform in
terms of the variation of the mechanical linkages, the kinematics error of the endeffector can be related to the manufacturing error of the linkages of both the planar
T
and vertical mechanisms L = [ l2 l4 l6 l8 k10 k12 ] by
E = EM L
49
(4.17)
e11
e
EM = 21
e31
e41
where
e31 =
8z
e12
e12
e13
e21
e32
e42
e22
e33
e43
e22
e34
e44
e23
e35
0
e11 =
l
,
2 ( d3 d1 )
e12 =
e21 =
l
,
4 y1
l
,
2 ( d5 d7 )
( x1 x2 ) l +
4 z ( d3 d1 )
e33 =
e11
e22 =
8z
( x2 x1 ) l
2
d5 d 7 )
(
2
l
( y1 y2 )
2
d3 d1 )
(
2
l
l
,
4 x2
e13 =
e23 =
e32 =
e13
e23
e35
k cos
2
( y2 y1 ) + ( x2 x1 )
k sin
2
( y2 y1 ) + ( x2 x1 )
( x1 x2 ) l
+
4 z ( d3 d1 )
( y2 y1 ) l ,
4 z ( d5 d7 )
e34 =
8z
e35 =
8z
( x2 x1 ) l
2
d5 d 7 )
(
2
l
( y y ) + ( x2 x1 )
2 k2 2 1
4
( y2 y1 ) l ,
4 z ( d5 d7 )
( x2 x1 )
l
( y2 y1 )
e42 =
2
( x2 x1 ) 2 + ( y2 y1 ) 2 ( d 3 d1 )
d3 d1 )
(
2
2 l
50
( x2 x1 )
l
( y2 y1 )
e41 =
2
2
2
( x2 x1 ) + ( y2 y1 ) ( d3 d1 )
d3 d1 )
(
2
2 l
4
( y1 y2 )
2
d3 d1 )
(
2
l
k
2
( y2 y1 )
x2 x1 )
(
l
e43 =
,
2
d5 d 7 )
( x2 x1 )2 + ( y2 y1 ) 2
(
d
d
(
)
7
2 l2 5
( y2 y1 )
x2 x1 )
(
l
e44 =
+
.
2
d
d
( x2 x1 )2 + ( y2 y1 )2
(
)
5
7
2 l 2 ( d5 d 7 )
4
The error manipulation matrix EM in (4.17) defines how the variations of the
mechanical linkages are transformed into kinematics error in the end-effector
coordinates G(x, y, z, ). In most manipulators, EM has a magnification effect so that
L is amplified to become larger errors at the end-effector. Since the size of L is

prescribed by the machining capability, the kinematics positioning accuracy can only
be improved by making EM as small as possible. In the next section, the design of
objective functions is considered to demonstrate how the parameters of the linkages
can be optimized to ensure that EM and hence E is small.
4.3 Design of Objective Functions

Let EM be factorized using Singular Value Decomposition as:
EM = USV T
(4.18)
where U = [u1u4] and V = [v1v4] are (part of) unitary matrices, and
S = diag ( 1 4 ) is a diagonal matrix with the singular values i (i = 1,2,3,4)

satisfying 1 4 0. The singular value decomposition can be regarded as a
principal component analysis of the error magnification matrix EM, with the singular
values representing the error amplification along different principal directions
defined by the singular vectors. From (4.18), since U and V are unitary, the error
magnification matrix can be reduced by minimizing the singular values of S.
51
From the optimization approach proposed by Ryu and Cha in [33], it is proposed that
the product of the singular values, (i.e. i) is minimized over a workspace volume.
The drawback of using the product of the singular values as an objective function in
the minimization is that very little control can be exercised over the magnitude of
individual singular values. For example, the product can be suppressed by making
only one particular singular value small while leaving the others unchecked. This
means that error amplification is reduced only along one direction of the workspace
corresponding to the suppressed singular value. It would be more desirable to reduce
all singular values of EM as a whole so that error amplification is suppressed
uniformly in all directions of the workspace.
We note that the Frobenius norm of EM is related to the singular values by:
EM
2
F
= S
2
F
= i2
(4.19)
i =1
Since all the singular values are bounded above by || EM || F, the minimization of the
Frobenius norm of EM ensures that all singular values are bounded below this value.
The minimization of two objective functions is proposed in the next section for the
purpose of suppressing the end-effector kinematics error over an operational
workspace W.
4.3.1 Minimizing the Integral of
EM
2
F
Workspace
Since EM depends not only on the kinematics parameters ( l , k ), but also on the
position (x, y, z, ) of the end-effector, it is necessary to ensure that EM remains small
over a chosen operational workspace W. For this purpose, we define an objective
function to be the integral of EM
2
F
over the operational workspace
f1 ( l , k ) =
1
W
i =1
52
2
i
dw
(4.20)
The kinematics parameters l and k are then determined by solving the minimization
problem
min f1 ( l , k )
l ,k
This means that l and k are optimized to reduce the average (or typical) error
amplification over W. The accuracy of the system is however often expressed in
terms of the worst-case error. For this reason, a second objective function for
reducing the worst-case kinematics errors magnification is considered.
4.3.2 Minimizing the Maximum of EM
2
F
Workspace
The objective function for the minimization of the worst-case kinematics error
magnification over the operational workspace W is defined as
f 2 ( l , k ) = max i2
( x , y , z , )W i =1
(4.21)
Let ( l*, k* ) be the solution to the minimization problem

min f 2 ( l , k )
l ,k
f2 ( l*, k* ) can be used to obtain a hard upper bound on the end-effector kinematics
error over the entire operational workspace W, which can then serve as a
specification of the system in the sense that all points within W will have an error no
larger than this value. The minimization of f2 ( l , k ) thus have the meaning of
designing the system for the best specification.
53
4.4 Optimization process

The objective of the optimization algorithm is to minimize the kinematics error of the
end-effector of the manipulator in the operational workspace. A cubic volume in the
centre of the manipulator workspace is selected to form the operational workspace.
The size of the operational workspace is equal to (x, y, z) = (15mm, 15mm,
2mm). The operational workspace is divided into 432 grid points for evaluating the
numerical integration in the case of the objective function f1 and for taking the
maximum in the case of f2.
The design variables of the optimization are the length l of the linkage of the two 2DOF planar mechanism and the length k of the linkage of the triangular structure as
shown in Figures 4.1 and 4.2. The maximum error budget of both l and k is set at
1m. The magnitude of l and k are constrained within the linkage boundaries to
avoid excessive geometrical error magnification. The boundary singularities of the
manipulator can be determined by the geometrical characteristic equations
k=
l=
( y2 y1 )
2
( d3 d1 )
2
and
x1 = x2
or
l=
(4.22)
( d5 d 7 )
2
(4.23)
These boundary singularities are avoided in the optimization process by constraining

the magnitude of l and k.
A constrained nonlinear optimization method is used for the minimization of the

kinematics error of the end-effector. The Sequential Quadratic Programming (SQP)
method is used for minimizing the value of the objective function. In each iteration,
the Hessian matrix of the Lagrangian function is calculated using the BFGS
(Broyden, Fletcher, Goldfrab and Shanno) method. The projected gradient is
calculated to determine the step size of the change of the design variables. Finally,
the design variables are modified for the next iteration. The iteration is terminated
54
when the difference of the objective function derivative in two iterations is less than
a certain minimum tolerance. In this case, the minimum tolerance of the objective
function derivative is defined to be 10-6.
4.5 Results
A given set of dimensional parameters of the linkages is used to generate the surface
of the proposed objective functions. We will compare the results of the proposed
objective function with that of [33], as there are no other existing method for
minimization of the kinematics error. The objective function proposed in [33] using
the product of singular values over the operational workspace W is defined as
f (l, k ) =
1
W
dw
i
W i =1
A plot of the surfaces of objective functions f( l , k ), f1 ( l , k ) and f2 ( l , k ) over the

feasible region of the constrained parameters l and k are shown in Figures 4.3-4.5.
Fig. 4.3 Surface of the objective function f( l , k )
55
Fig. 4.4 Surface of the objective function f1 ( l , k )
Fig. 4.5 Surface of the objective function f2 ( l , k )
56
From the surface plots of the objective functions f( l , k ), f1 ( l , k ) and f2 ( l , k ) shown in

Figures 4.3-4.5, no local minimum of the objective functions exists within the
feasible region. This means that under the constraints imposed on the kinematic
parameters, the optimal solution lies on the boundary of the feasible region.
Figures 4.6 and 4.7 show the plot of the values of EM
2
F
at the 432 individual test
points in the operational workspace W. These plots show how the objective functions
are related to optimal value in the two different schemes of optimization.
Fig. 4.6 Values of EM
2
F
at the individual test points in the workspace W in the case

when f1 ( l , k ) is optimized
57
Fig.4.7 Values of EM
2
F
at the individual test points in the workspace W in the case

when f2 ( l , k ) is optimized
From Figures 4.6 and 4.7, it can be observed that although the optimal value of f1
(=3.535) is smaller than that of f2 (=3.70), the latter represents a hard upper bound
for all grid points in the workspace whereas f1 can only be read as an average error.
Furthermore, the range as well as the scattering of the points is slightly smaller in the
case of the optimal solution for f2 .
The initial parameters of the linkages l and k, and the optimal parameters for the
objective functions f, f1 and f2 after optimization are shown in Table 4.1.
l (mm)
k (mm)
Objective function
Initial parameters of f(l,k)
130
130
0.562
Initial parameters of f1(l,k)
130
130
153.125
Initial parameters of f2(l,k)
130
130
626.627
Optimal solution for f(l,k)
70
119.3
0.0293
Optimal solution for f1(l,k)
70
118.6
3.535
Optimal solution for f2(l,k)
70
116.6
3.70
58
Table 4.1 Optimization results of the linkages and the objective functions of the
manipulator
x (m)
y (m)
z (m)
(mdeg)
Minimum
3.79
3.79
-91.66
-602.59
Maximum
106.09
106.09
-1.52
602.59
Average
12.17
12.17
-8.86
Standard deviation
15.89
15.89
13.89
139.98
Table 4.2 Statistical analysis of the initial kinematics positioning accuracy
Table 4.2 shows that with the initial kinematics parameters (i.e., without performing
any kinematics optimization), a variation of only 1m in each of the linkage of l and
k, the average kinematics positioning error of the end-effector is magnified to 12m
in the XY plane and 8m in the Z-axis.
The kinematics positioning error after optimization using the objective functions
f( l , k ), f1 ( l , k ) and f2 ( l , k ) are shown in Tables 4.3-4.5 respectively.
x (m)
y (m)
z (m)
(mdeg)
Minimum
0.69
0.69
0.67
-2.8
Maximum
1.05
1.05
0.89
2.8
Average
0.84
0.84
0.8
Standard deviation
0.093
0.093
0.052
1.25
Table 4.3 Statistical analysis of the kinematics positioning accuracy optimized by f
x (m)
y (m)
z (m)
(mdeg)
Minimum
0.68
0.68
0.68
-2.72
Maximum
1.03
1.03
0.89
2.72
Average
0.83
0.83
0.8
Standard deviation
0.089
0.089
0.049
1.225
59
Table 4.4 Statistical analysis of the kinematics positioning accuracy optimized by f1
x (m)
y (m)
z (m)
(mdeg)
Minimum
0.65
0.65
0.73
-2.47
Maximum
0.96
0.96
0.9
2.47
Average
0.78
0.78
0.83
Standard deviation
0.079
0.079
0.042
1.147
Table 4.5 Statistical analysis of the kinematics positioning accuracy optimized by f2

Tables 4.3-4.5 show that a significant reduction of the kinematics positioning error of
the end-effector has been achieved by minimizing either f1 ( l , k ) or f2 ( l , k ). With the
optimized solution, the kinematics error of the end-effector is of the same order of
magnitude as the variation of the linkages l and k. This is in contrast to an error
magnification that is usually found in many mechanisms. Indeed, if the initial
parameters of the length of the mechanical linkages are used in the present
mechanism, the variations in the linkages would result in a kinematics error of about
an order of magnitude larger than the errors in the linkages.
4.6 Summary
An optimization method is developed for suppressing the kinematics error
magnification of the parallel manipulator. The error manipulation matrix is derived
for the transformation of the mechanical linkages variation to the end-effector
kinematics error. Two objective functions are developed for characterizing the error
magnification effect and the optimal kinematics parameters are determined by
minimizing these functions. A significant improvement of the kinematics positioning
accuracy has been achieved, which demonstrates the usefulness of the proposed
optimization method for high accuracy semiconductor packaging applications. The
objective functions developed in this thesis have been found to give better results
compared with the objective function proposed in [33].
Furthermore, the error
kinematics analysis for the 4-DOF parallel manipulator, together with suitable
choices of the kinematic parameters obtained using the optimization approach
60
developed in this thesis, shows that it is possible to achieve a high positioning

accuracy which matches the error tolerence attainable in the manufacturing of the
linkages. The optimization results provide a justification for the structural design of
the 4-DOF manipulator in terms of its ability to suppress kinematics error.
61
Chapter 5
Dynamic Modelling of the Parallel
Manipulator
5.1 Introduction
Dynamic modelling of parallel mechanisms is in general difficult and tedious due to
the closed-chain kinematics structure leading to highly coupled dynamics, but is
essential for the motion analysis of parallel manipulators under a given force profile
and for estimating the motion capability in terms of profile tracking and the required
driving forces. Several dynamic modelling approaches including Lagrangian
formulation [34], Euler-Lagrangian approach [35], the Hamilton Principle [36], [37],
the Principle of Virtual Work [38], [39] and Newton-Euler approach [40], [41], [42]
are employed to develop the mathematical model of parallel manipulators. In this
chapter, the dynamic modelling of the 4-DOF parallel manipulator is developed
using Recursive Newton-Euler method. In Recursive Newton-Euler method, the joint
velocity, acceleration and forces acting on the individual linkage can be obtained and
are useful for sizing the mechanisms and desired actuators during the design stage.
Since the kinematics design of the parallel manipulator has a nested structure, this
enables the dynamic modelling to be performed by a nested approach. The problem
of modelling the parallel mechanism is decomposed into that of modelling triangularmechanisms (T-mechanisms) first at the actuator level in the horizontal plane and
then at the end-effector level in the vertical plane. This approach affords a high
62
degree of decoupling between the triangular mechanisms in the modelling process.

The necessity of modeling the dynamics of the belt mechanism used for constraining
the motion of a link is also considered. Such constraint belts are commonly used in
high precision motion systems. It is subsequently shown by simulation that the effect
due to the elasticity of the belt is so small that the dynamics of the constraint belts
can be neglected in the manipulator model. With the benefits of the high-strength
material of the mechanism and low friction ball bearings, the joint elasticity and the
friction of the manipulator can also be neglected. Simulation results based on the
dynamic model are given to demonstrate the effectiveness of the model.
5.2 Dynamic Modelling using Recursive NewtonEuler Method

The dynamic model of a parallel manipulator can be expressed in the form [74], [76]:
F = M ( q ) q + C ( q, q ) + G ( q ) + Fd
(5.1)
where F denotes the actuating force vector, q represents the Generalized coordinates
vector, M represents the mass matrix of the manipulator, C represents the velocity
coupling vector, G is the gravitational force vector, and Fd is the external force
vector (if any).
The dynamical model of the parallel manipulator is derived in the following manner.
For each T-mechanism, the closed kinematics chain will be decomposed into two
serial kinematics chains with interaction between the two serial chains represented by
action and reaction forces at the ends of the serial chains. Recursive Newton-Euler
method will be used to derive a dynamic model for each serial chain and the action
and reaction forces will then be eliminated to give a model for the T-mechanism.
Given a kinematics chain, coordinate frames are defined using the DH-convention at
the outer joint of each link. The notation of [77] is adopted in the following Newton-
63
Euler formulation. A leading superscript i attached to a vector quantity indicates that

the vector is expressed in frame i.
Consider the ith link of a kinematics chain. The inertia force i f i * and moment i ni*
exerted at the centre of mass of link i expressed in frame i are given by [77]:
f i * = mi i vci
(5.2)
i *
i
n = i I i i i ii i I i ii
(5.3)
where mi is the mass of link i, i vci is the absolute linear acceleration of the centre of
mass of link i, ii is the absolute angular velocity of link i, and i I i is the inertia
matrix of link i about its centre of mass, all expressed in frame i.
Let i1Ri denotes the matrix representing the rotation from frame (i1) to frame i, and
i1
fi,i-1 and i1ni,i-1 be the reaction force and moment, respectively, exerted on link i by
link (i1) expressed in frame (i1), given by
i 1
fi ,i 1 =
i 1
Ri
i 1
ni ,i 1 =
i 1
Ri i ni +1,i + i ri + i rci i f i ,i 1 i rci i f i +1,i i ni*
f i +1,i mi i g i f i*
(5.4)
(5.5)
where i ri is the vector from the origin of frame (i1) to the origin of frame i, and i rci
is the vector from the origin of frame i to the centre of mass of link i, both expressed
in frame i.
5.3 Dynamic Modelling of the Planar Mechanism

Suppose the T-mechanism on the XY-plane is decomposed into two geometrically
symmetric serial limbs as shown in Fig. 5.1.
64
Fig. 5.1 Free-body diagram of the planar mechanism on the XY-plane
Since joint 1 is prismatic with joint axis along the X-axis, we have:
1 = 0,
1 = 0
vc1 = d1 0 0
and
and
For link 2, we have:
2 = 0 0 2
vc 2 = c 2 d1 r22 2
2 = 0 0 2
T
r22 s 2 d1 0
Let p and p be the interaction forces in the Y- and X-directions (respectively) acting
between the two serial limbs at the joint connecting links 2 and 4, and let p and p
be external forces (due to the vertical mechanism) acting at the same point. We may
assume that forces acting on limb 1 at the end of link 2 are ( p + p ) and p ,
while that on limb 2 at the end of link 4 are p and ( p + p ) . Using Newton-Euler
formulation, the forces and moments acting at the joints of the links can be obtained:
F1 = ( m1 + m2 ) d1 m2 r2c 222 m2 s 2 r22 + p
65
(5.6)
2 = l2 ( p + p ) c 2 l2 p s 2 + m2 r222 m2 r2 s 2d1 + I 22 = 0
(5.7)
F3 = ( m3 + m4 ) d3 m4 r4c 442 m4 s 4 r44 + ( p + p )
(5.8)
4 = l4 p c 4 + l4 ( p + p ) s 4 m4 r424 + m4 r4 s 4d3 + I 44 = 0
(5.9)
where we have set 2 and 4 to zero because the revolute joints at links 2 and 4 are
passive. From (5.7) and (5.9), we can solve for the interaction forces p and p in
terms of p and p . Substituting for p in (5.6) and making use of the symmetry
2 = 4 and l2 = l4= l, the force acting on link 1 can be expressed as
mr
mr
F1 = m1 + m2 2 2 d1 + 4 4 d3 m2 c 2 r222 +
2l2
2l2
m4 r42 + I 4 + m2 r22 + I 2
p c 2 p
m2 s 2 r2 2 +
2l2 s 2
2 s 2
2
(5.10)
The joint angle 2 in (5.10) can be eliminated by expressing it in terms of the

positions of the linear actuators:
2 = 4 = cos 1 ( ( d 3 d1 ) 2l ) .
Applying a similar argument to the force balance for each of the four linear actuators,
the dynamics of the two structures on the horizontal plane can be expressed as
Fp = M p ( q ) q + C p ( q, q ) + D p Fdv ,
where
Fp = [ F1
F3
F5
66
F7 ] ,
(5.11)
p q
Fdv = p
q = [ d1
d3
d5
q ,
d7 ] .
Explicit expressions for the matrices Mp , Cp and Dp in (5.11) are given in the
Appendix (Section 5.8.1). From the form of Mp , Cp and Dp , the two horizontal Tmechanisms appear to be decoupled from each other, which would be the case if not
for the vertical T-mechanism straddling them. In the next section, the dynamic model
of the vertical T-mechanism which interacts with the horizontal structures through
Fdv is determined.
5.4 Dynamic Modelling of the Vertical Mechanism

To model the dynamics of the vertical T-mechanism, we decompose it into two serial
limbs, as shown in the free-body diagram of Fig. 5.2.
Fig. 5.2 Free-body diagram of the vertical T-mechanism

67
The coordinate frames of the vertical T-mechanism are assigned according to the DH
convention as shown in Fig. 5.3.
Fig. 5.3 Coordinate frame of the vertical T-mechanism
In Fig. 5.2, the block on the horizontal plane with mass mp represent two prismatic
joints driven by forces p and p , and are constrained to have translational motion
in the x1 and y1 directions only. Therefore, the rotational velocities 11 and 22 , and
hence the rotational accelerations 11 and 2 2 of the first two links are zero. The
linear accelerations can be expressed as
vc1 = [0
y1 0]
and
68
vc 2 = [
y1

x1
0] .
Link 9 with mass m9 has a revolute joint for rotational motion of the vertical motion
mechanism along the Z-axis. The rotational velocity 33 and the rotational
acceleration 3 3 are
3 = 0 9 0
and
T
3 = 0 9 0 .
The linear acceleration at the centre of mass of the link 9 is given by
vc 3 = [ c 9
y1 s 9
x1 0 s 9
y1 + c 9
x1 ] .
Link 10 is supported on link 9 by a revolute joint with a horizontal joint axis. The
rotational velocity 4 4 and the rotational acceleration 4 4 of link 4 are
4 = s109 c109 10 ,
4 = c10910 + s109 s10910 + c109 10 .
The linear acceleration at the centre of gravity of link 10 is:
(
(
c10 ( c9
y1 + s9
x1 ) r10 c 21092 + 102
4
vc 4 = s10 ( c9
y1 + s9
x1 ) + r10 c10 s1092 + 10
s

9 y1 c 9 x1 + r10 2 s10 910 c10 9
Let , and be the interaction forces (with along the z4-axis, perpendicular
to on a horizontal plane, and vertical), and be the moment about a vertical
axis acting between the two serial limbs at the joint connecting links 10 and 12, and
let , and be external forces (e.g. due to inertia load at the end effector) and
be the external moment acting at the joint. Hence, if we assume that forces acting
on link 12 are , , and the moment is , then the forces acting on link 10
are ( + ) , ( + ) , ( + ) and the moment is ( + ) . Using Newton-Euler
69
formulation in (5.2)-(5.5), the forces driving the small platform at the joint
connecting links 2 and 4 are:
p = ( m p + m9 + m10 ) y1 m10 r10 s 9 c109 m10 r10 c 9 s1010 m10 r10 c 9c1092
+ 2m10 r10 s 9 s10910 m10 r10 c 9 c10102 + ( + ) c 9 + ( + ) s 9 ,
p = ( m p + m9 + m10 )
x1 + m10 r10c 9c109 m10 r10 s 9 s1010 m10 r10 s 9c1092
2m10 r10c 9 s10910 m10 r10 s 9c10102 + ( + ) s 9 ( + ) c 9 ,
(5.12)
(5.13)
and the moments acting on the vertical links 9 and 10 are determined to be:
9 = ( I 9 + I10 )9 m10 r10 s 9 c10 y1 m10 r10 c 9 c10 x1 + m10 r10 2c 2109
(5.14)
2
2m10 r10 c10 s10910 k10 c10 ( + ) + ( + ) ,
2
10 = m10 r10 c10 g m10 r10c10 s10 y1 + m10 r10 s 9 s10
x1 + m10 r10 + I10 10
2
+ m10 r10 c10 s1092 k10 s10 ( + ) + k10 c10 ( + ) .
(5.15)
By an argument based on geometric symmetry, 11 and 12 are given by:
2
11 = ( I11 + I12 )11 m12 r12 s11c12 y2 m12 r12 c11c12
x2 m12 r12 c 21211
2
+ 2m12 r12 c12 s121112 k12c12 ,
2
12 = m12 r12 c12 g + m12 r12 c11s12 y2 + m12 r12 s11s12
x2 m12 h12 + I12 12
(5.16)
(5.17)
m12 r12 c12 s12112 + k12 s12 + k12 c12 .
Since the revolute joints for the links 9, 10, 11 and 12 are all passive, we may equate
(5.14), (5.15), (5.16) and (5.17) to zero. These four equations can be solved for , ,
and , giving (only , are needed in the sequel and stated here):
70
= m 10 ( s 9
x1 c 9
y1 ) m 12 ( s11
x2 + c11
y2 )
+ (m 10 + m 12 )
(5.18)

c10
c10
10 + (m 10 r10 + m 12 r12 )c1092 +
g+
,
s10
s10
2 2 s10
= m 10 (c9 x1 + s9 y1 ) m 12 (c11x2 s11 y2 )
(5.19)
I

+ (m 10r10 m 12 r12 )c10 +
.
9 + 2(m 10r10 m 12 r12 )s10910 +
2k10c10
2 2k10c10
where
I = I 9 + I10 I11 I12 ,
m i =
mi ri
2ki
( i = 10, 12) and = m 10 r10 + m 12 r12 +
I10 + I12
2k10
By (5.18) and (5.19), we can replace and in (5.12) and (5.13) to give the
dynamical equations relating x1 and y1 to p and p . Using a similar argument, we
can obtain the equations relating x2 and y2 to q and q . Note that the variables 9,
10, 11 and 12 can be expressed in terms of x1, y1, x2 and y2 as
9 = 11 = tan 1 ( x2 x1 y2 y1 ) and 10 = 12 = cos 1
( x1 x2 ) + ( y2 y1 )
2k .
Allowing 9 and 10 to stay as redundant variables for now, the dynamics of the
vertical motion mechanism can be expressed as:
Fdv = M v q + Cv ( q, q ) + Gv ( q ) + Dv Fef ,
where
and
q = [ x1
y1
x2
(5.20)
y2 9 10 ] ,
] [
T
T
Fef = = mef z mef x mef y I ef .
(5.21)
Fef represents the inertia reactions in the x, y, z and directions due to the mass mef
and moment of inertia Ief of the end-effector. The expressions for the matrices Mv , Cv,
71
Gv and Dv are given in the Appendix (Section 5.8.2). Equations (5.11), (5.20) and
(5.21) now provide a complete dynamic model for the parallel manipulator.
5.5 Modelling of the Belt Mechanism

The small platforms at the end of the two planar T-mechanisms are constrained by
belt mechanisms to have only translation but not rotational motion. Despite the use
of high-stiffness belts, the elastic property of the belts may lower the position
accuracy of the end effector. To assess the effect of the belt elasticity, we may
include the dynamics of the belt mechanism into the dynamic modelling. In practice,
the axes of links 9 and 11 of the vertical T-mechanism are supported on points on the
small platforms at offsets of dy and dx from the axis of the belt mechanism. Due to
the elasticity of the belt, the small platforms may in fact undergo very slight rotations
p and q. To model such effects, the small platforms supporting the vertical
mechanism will be regarded as links of lengths dy and dx jointed to the end of the
planar T-mechanisms by revolute joints with joint variables p and q and driven by
torques p and q, respectively. Fig. 5.4 shows the addition of such a link to the end
of the T-mechanism that is driven by p and p (where apart from the extra link, the
other joints and links are labelled as in Fig. 5.2).
72
Fig. 5.4 Free-body diagram of the vertical T-mechanism with belt mechanism
Similarly, a link with joint variable q driven by torque q can be added to the end of
the T-mechanism driven by q and q . These extra links can be incorporated into
the dynamic modelling given in Section 5.4. The torques p and q represent the
restoring forces due to the belt mechanisms which are supposed to constraint the
small platforms at their nominal positions p= q = 0. The model given in Section 5.4
in fact assumes that the elastic effects of the belts are negligible and the condition
p= q = 0 is exactly satisfied.
To assess the effects of the elasticity of the belts, we have constructed a full model
incorporating the two extra links as described above with additional revolute joint
variables p and q. In this full model, the driving torques p and q are
p = kbp and q = kb q
where kb is the stiffness of the belts. The full model is too lengthy to be given here.
However, simulation results using the full model will be presented in the next section
73
to show that for a sufficiently large kb (which is the case by design) p and q are
very small, thus justifying the use of the model developed in Section 5.4.
5.6 Dynamic Simulations

The dynamic model derived in Sections 5.3 and 5.4 consisting of (5.11), (5.20) and
(5.21) can be readily implemented using MATLAB and SIMULINK for simulation
purpose. Fig. 5.5 shows a block diagram of the model of the manipulator, where
(5.11) is represented as
q = M p 1 ( q ) Fp C p ( q, q ) D p Fdv
and Fdv is computed using (5.20).
Fig. 5.5 Block diagram of the dynamic model of the 4-DOF parallel manipulator
For the purpose of simulating the response of the dynamic model, the linear actuators
are driven using sinusoidal force profiles of the form:
74
F1 = sin t ,
F3 = sin ( t )
(5.22)
F5 = sin t ,
F7 = sin ( t )
(5.23)
where = 10.5 N is a force constant, and the phase angle is introduced so that the
motion of the linear actuators are not synchronized. Two different frequencies, =
89.8 and 125.7 rad/s corresponding to periods of 70 and 50 ms respectively, are used
to test the position tracking response of the prototype. A simulation is performed for
each of the following cases:
o
(i) (a) ( , ) = (89.8, 30 ),
(b) ( , ) = (89.8, 330 )

o
(ii) (a) ( , ) = (125.7, 150 ),
(b) ( , ) = (125.7, 210 )
(5.24)
(5.25)
The results of the simulations for the four cases listed in (5.24) and (5.25) are shown
in Figures 5.6-5.9. In each case the manipulator is started with initial conditions:
q = [ d1
d3
d5
d 7 ] = [ 79.5 139.5 149.5 89.5]
and
q = 0
Table 5.1 lists the dimensions of the linkages of the manipulator used for the
dynamic simulation.
Parameter
Value
Length (mm)
l2, l4, l6 and l8
70
r2 and r6
43
r4 and r8
46
k10 and k12
116.6
r10
64.35
r12
69.77
dx, dy and dz
10
Mass (kg)
75
m1, m3, m5 and m7
0.7
m2 and m6
0.104
m4 and m8
0.094
mp and mq
0.034
m9 and m11
0.019
m10
0.147
m12
0.13
End-effector mef
0.027
Moment of inertia (kgm2)

I2 and I6
7.2x10-5
I4 and I8
5.5x10-5
I9 and I11
1.86x10-6
I10
2.39x10-4
I12
1.86x10-4
End-effector Ief
7.62x10-6
Table 5.1 Kinematics and dynamic properties of the 4-DOF parallel manipulator for
simulation
Since the dynamic models are based on the linear actuator positions d1, d3, d5 and d7,
the end-effector trajectories has to be determined using the forward kinematics
equations given in Chapter 2. The end-effector trajectories computed from the
simulated actuator positions are shown in Fig. 5.10.
76
Fig. 5.6 Simulated actuator positions for case (i)(a) ( , ) = (89.8, 30 )
Fig. 5.7 Simulated actuator positions for case (i)(b) ( , ) = (89.8, 330 )
77
Fig. 5.8 Simulated actuator positions for case (ii)(a) ( , ) = (125.7, 150 )
Fig. 5.9 Simulated actuator positions for case (ii)(b) ( , ) = (125.7, 210 )
78
Fig. 5.10 Simulated trajectories of the end-effector
To assess the effect of the elasticity of the belt mechanism on the accuracy of the
system, simulations are performed using the full model described in Section 5.5. Fig.
5.11 shows the response of p and q for the same input force profiles as given in
case (i)(a) above.
o
It can be seen that the magnitudes of p and q are within 0.03 under dynamic
conditions, and become negligible under steady-state conditions. As it is the static
accuracy at the end of a trajectory that is important in manufacturing placement
operations, this shows that for practical purposes, the elasticity of the belt mechanism
can be neglected.
79
Fig. 5.11 Simulation of the response of p and q with force inputs of case (i)(a)
5.7 Summary
The dynamic model of the proposed 4-DOF parallel manipulator is developed in this
chapter. The derivation of the mathematical model of the manipulator is simplified
by modelling the multiple closed kinematics chains separately, and by decomposing
each closed T-mechanism into two open kinematics chains with interaction forces. A
complete dynamic model for the parallel manipulator is derived and demonstrated
how the elasticity of the constraint belt mechanism can be taken into account in a full
dynamic model. From the simulations results, the motion capability of the parallel
manipulator under a given force profile can be estimated. Simulation results
including a dynamic model of the belt mechanism demonstrated that the belt
elasticity has negligible effect on the end-point accuracy particularly under steadystate conditions. The dynamic model obtained in this chapter provides a basis for
development of controllers for the parallel manipulator for high-precision positioning
to be considered in Chapter 7.
80
5.8 Appendix
5.8.1 Expressions for matrices in the dynamic model (5.11):
m2 r2
m4 r4 + m2 r2
a
0
0
m12 l + a
2l2
2
m4 r4 + m2 r2 + a m + m4 r4 a
0
0
34
2l2
l2
Mp =
mr
m r + m6 r6
0
0
m56 + 6 6 b 8 8
+ b
l6
2l6
m8 r8 + m6 r6
m8 r8
b
+b
0
0
m78
2l6
l6
where
a=
m12 = m1 + m2 ; m34 = m3 + m4 ; m56 = m5 + m6 ; m78 = m7 + m8
m4 r42 + I 4 + m2 r22 + I 2
m8 r82 + I8 + m6 r62 + I 6
;
b
=
4l22 s 2 2
4l62 s 2 6
1
Cp =
2
a c 2
d1 d 3
2
l2 s 2
D p = diag
c 2
2 s
2
2
2 s 2
1
2
1
2
a c 2
d1 d 3
l2 s 2 2
1
2
,
1
2
c8
2 s 8
c 8
2 s 8
b c 6
d 7 d5
l6 s 2 6
b c 6 2
d7 d5
l6 s 2 6
5.8.2 Expressions for matrices in the dynamic model (5.20):

0
M
M v = v 21
M v 31
M v 41
M v12
M v13
M v14
M v15
0
M v 32
M v 42
M v 23
0
M v 43
M v 24
M v 34
0
M v 25
M v 35
M v 45
81
M v16
M v 26
M v 36
M v 46
M v12 = M v 21 = ( m p + m9 + m10 m 10 )
M v13 = M v 24 = 2m 12 s 9 c 9
M v14 = M v 23 = m 12 ( s 2 9 c 2 9 )
M v15 =
I s 9
+ ( m 10 r10 m10 r10 m 12 r12 ) s 9 c10
2k10 s10
M v16 =
M v 25 =
c 9
m10 r10 c 9 s10
s10
I c 9
( m 10 r10 m10 r10 m 12 r12 ) c 9c10
2k10 s10
M v 26 =
s 9
m10 r10 s 9 s10
s10
M v 31 = M v 42 = 2m 10 s 9 c 9 ; M v 32 = M v 41 = m 10 ( s 2 9 c 2 9 )
M v 34 = M v 43 = ( mq + m11 + m12 m 12 )
M v 35 =
I s 9
+ ( m 12 r12 m12 r12 m 10 r10 ) s 9 c10
2k10 c10
M v 36 =
M v 45 =
c 9
m12 r12c 9 s10
s10
I c 9
+ ( m 10 r10 m 12 r12 + m12 r12 ) c 9 c10
2k10c10
82
M v 46 =
s 9
m12 r12 s 9 s10
s10
( m 10 r10 + m 12 r12 m10 r10 ) c 9c1092 m10 r10c 9 c10102 + 2 ( m 12 r12 m 10 r10 + m10 r10 ) s 9 s10910
( m r + m 12 r12 m10 r10 ) s 9c1092 m10 r10 s 9c10102 2 ( m 12 r12 m 10 r10 + m10 r10 ) c 9 s10910
Cv = 10 10
( m 10 r10 + m 12 r12 m12 r12 ) c 9c1092 m12 r12c 9c10102 2 ( m 12 r12 m 10 r10 + m10 r10 ) s 9 s10910
2
2

( m 10 r10 + m 12 r12 m12 r12 ) s 9c10 9 m12 r12 s 9 s1010 + 2 ( m 12 r12 m 10 r10 + m10 r10 ) c 9 s10 910
Gv = [G1 G2
G1 G2 ]
where G1 =
s9
2
c9
2
Dv =
s9
2
c9
2
( m 10 + m 12 ) c9c10 , G
s10
c9
2
c9c10
2 s10
s9
2
s9c10
2 s10
c 9
2
c9c10
2 s10
s9
2
s9c10
2 s10
s 9
2k10c10
c 9
2k10c10
s9
2k10c10
c9
2k10c10
83
( m 10 + m 12 ) s9c10
s10
Chapter 6
Prototype Design
6.1 Introduction
In this chapter, a prototype of the 4-DOF parallel manipulator is developed based on
the results of the kinematics analysis of Chapter 4. We will described how the design
methodology to realize the parallel manipulator as a practical manipulation platform.
The design of the single degree-of-freedom mechanical joints and the belt
mechanism are first introduced to show the approach for achieving high kinematics
accuracy. Linear AC servo motors and high precision linear encoders are introduced
to show the benefits of the direct-drive linear motion systems to be used on the
parallel manipulator. The motion control platform of the parallel manipulator
prototype, including the Pulse Width Modulation (PWM) AC servo amplifier and
PC-DSP system are described. The current control loop of the servo amplifier is
verified using frequency domain approach. This ensures that the servo amplifier has
an adequate cut-off frequency for good tracking response at low frequency and
minimum phase shift at high frequency, which help to reduce the current ripple and
noise within the current feedback loop. The accuracy of the mathematical model of
the 4-DOF parallel manipulator is assessed by comparing the dynamic response of
the manipulator prototype under open-loop control with the simulation results of
Chapter 5.
84
6.2 Mechanical
Design
of
the
Manipulator
Prototype
A prototype of the 4-DOF parallel manipulator has been designed and constructed as
shown in Fig. 6.1 with target applications for semiconductor packaging. The length
of the mechanical linkages used on the parallel manipulator are designed and
optimized using the approach developed in Chapter 4. The linkages and the joints of
the manipulator are fabricated using Aluminum alloy AL-6061 with the advantages
of high rigidity and low moving mass due to the mechanical parts of the manipulator.
Fig. 6.1 Prototype of the 4-DOF parallel manipulator
In the prototype, four ironless linear AC servo motors are used as actuators to drive
the manipulator and installed on a common rigid base such that the payload of the
moving parts of the manipulator is even-distributed to the actuators. With the
benefits of the direct-drive approach using linear servo motors, the energy loss and
kinematics error due to traditional rotary transmission system such as ball-screw or
gearbox can be avoided. The three-phase motor coil and the double-layer permanent
85
magnet track of a linear motor are shown in Fig. 6.2. High precision optical linear
encoders with 0.2m/count resolution are used to measure the actual positions of the
prismatic joints. The glass scale and the read head of the linear encoder are installed
on the rigid mount as shown in Fig. 6.3. The moving part of the linear motor travels
on the linear motion guide to provide a translational motion for the prismatic joints.
Mechanical stoppers are installed on the linear motion guide to limit the travel range
of the linear motors as stipulated in condition (2.4).
Three-phase
motor coil
Permanent
Magnet Track
Fig. 6.2 Linear AC servo motor of the parallel manipulator
86
Read-head
Glass scale
Fig. 6.3 High precision linear encoder system
All of the high precision passive rotational joints in the parallel manipulator are
designed using double-shield angular contact ball bearings pair with wave spring
preload. A typical rotational joint design of the linkage l2 on the prismatic joint is
shown in Fig. 6.4. In the rotational joint, the lower ball bearing is supported by a
rigid locking mechanism to form the base of the bearing pair. A spacer is installed
between the lower and upper ball bearings to avoid direct contact of the roller balls in
the bearing during the rotational motion. A wave spring with bearing cover are
mounted on top of the upper bearing to provide a mechanical preload to the whole
bearing pair. From the benefits of the angular contact bearing pairs and the wave
spring preload, the tolerance gap inside the moving part of the joint can be
minimized and the slipping motion between bearing pair can be avoided. Hence, the
accuracy of the proposed passive rotational joint design can be achieved.
87
Doubleshield ball
bearings
Bearing
covers
Wave spring
preload
Spacer
Rigid
locking
mechanism
Fig. 6.4 Design of high precision passive rotational joint
The belt mechanism constraining the orientation of the small platform at the outer
joints of the planar T-mechanism is shown in Fig. 6.5. The high stiffness of the belt
provides adequate torque to constrain the small platform to have only translation but
not rotational motion. The end-effector of the parallel manipulator is mounted on the
vertical T-mechanism as shown in Fig. 6.6.
Fig. 6.5 Belt mechanism of the parallel manipulator

88
Fig. 6.6 End-effector of the parallel manipulator
The mechanical design, determination of the moving mass and moment of inertia of
the linkage are performed and verified with the CAD software [78]. Table 6.1 lists
the dimensions and mechanical characteristics of the linkages of the manipulator,
based on which a prototype has been constructed.
Parameter
Value
Actuator ranges (mm)

d1min , d7min
59.5
d1max , d7max
99.5
d3min , d5min
119.5
d3max , d5max
159.5
Length (mm)
l2, l4, l6 and l8
70
r2 and r6
43
r4 and r8
46
k10 and k12
116.6
r10
64.35
89
r12
69.77
dx, dy and dz
10
Mass (kg)
m1, m3, m5 and m7
0.7
m2 and m6
0.104
m4 and m8
0.094
mp and mq
0.034
m9 and m11
0.019
m10
0.147
m12
0.13
End-effector mef
0.027
BGA substrate
0.0023
Moment of inertia (kgm2)

I2 and I6
7.2x10-5
I4 and I8
5.5x10-5
I9 and I11
1.86x10-6
I10
2.39x10-4
I12
1.86x10-4
End-effector Ief
7.62x10-6
Table 6.1 Kinematics and dynamic properties of the parallel manipulator
6.3 Electronic Design of the Manipulator Prototype

The motion control platform of the parallel manipulator is developed using a PCDSP controller with four PWM AC servo amplifiers. The block diagram of the
motion control platform is shown in Fig. 6.7.
90
Fig. 6.7 Block diagram of the motion control platform of the parallel manipulator
In Fig. 6.7, the functions of the host PC include determination of forward and inverse
kinematics, profile trajectory planning and data logging of the motion performance
of the parallel manipulator. The DSP servo controller is used to implement and
performed the closed-loop control of the manipulator using position feedback from
the linear encoders. In the DSP servo controller, the controller output Ui(t) of the
linear motor i is converted to a two phase control voltage Uai(t) and Ubi(t) expressed
as
U ai ( t ) = U i ( t ) cos ei ( t ) +
2
and
U bi ( t ) = U i ( t ) cos ei ( t ) +
where ei ( t ) is the electrical angle determined by the position of the linear motor
and the magnetic pole-pitch. The servo amplifier then converts the control voltages
to three-phase currents Iai(t), Ibi(t) and Ici(t) driving the linear motor as
I ai ( t ) = K c U ai ( t ) ,
I bi ( t ) = K c U bi ( t )
and
I ci ( t ) = ( I ai ( t ) + I bi ( t ) )
where Kc is the current gain of the servo amplifier.

An analogue PI feedback current control system is designed using the electrical
characteristics of the linear motor given in table 6.2.
91
Parameter
Value
Phase-to-phase resistance ()
0.96
Phase-to-phase inductance (mH)
0.36
Peak Current (A)
12.5
RMS Current (A)
3.0
Voltage-Current Gain (A/V)
1.25
DC Buss voltage of the servo amplifier (V)
150
Coil connection type
Star
Table 6.2 Electrical characteristics of the linear AC servo motor
The current feedback controller is implemented in the servo amplifier using the
current sensors in the PWM stage to acquire the current feedback and control the
tracking of the output current as shown in Fig. 6.8.
Fig. 6.8 Block diagram of the current loop control of the servo amplifier
In the closed-loop control of the parallel manipulator, the current control bandwidth
is defined as the cut-off frequency at which the gain magnitude is -3dB below the
initial gain acquired at the frequency of 100 Hz. It is necessary to design the current
controller to provide minimum gain variation and phase shift at frequencies below 1
KHz with adequate current control bandwidth for the current tracking response of the
amplifier. Fig. 6.9 shows an analogue PI controller with a resistor-capacitor pair that
can be designed using the pole-placement approach [79].
92
Fig. 6.9 Schematic diagram of the analogue PI current controller
In Fig. 6.9, the impedances of the PI controller Z and the servo motor Zm are
expressed as
Z = R+
1 1 + sRC
=
sC
sC
and
Z m = Rm + sLm = 1 + s
Lm
Rm
where R is the proportional gain, C is the integral gain. In Fig. 6.9, Vin is the
command input voltage, Vo is the output voltage, Vf is the feedback voltage, I is the
driving current of the servo motor with phase-to-phase resistance Rm and phase-tophase inductance Lm, Rc and Rf are the command resistance and feedback resistance
of the analogue PI controller respectively. Let Rc = Rf, the driving current I of the
servo motor can be expressed as
I = Vin
1 + sRC

Lm
Rm 1 + s
Rm

(1 + sRC )
+
sRc C
(6.1)
Using the pole-placement approach with C equal to 22nF, the proportional gain R
can be determined from the expression of
93
Lm
= RC
Rm
(6.2)
Substituting (6.2) into (6.1), the driving current of the servo motor can be simplified
to a first order linear system as
1
I = Vin
sRm Rc C
(6.3)
From (6.3), the current control loop can provide the required tracking response
similar to that of a first order system.
The response of the current control loop of the servo amplifier is verified using the
swept-sine identification method. Sinusoidal control signals with the peak amplitude
of 1V (1.25A) over the frequency range of 100 Hz to 10 KHz is generated as input to
the servo amplifier to produce driving current for the linear motor. With these input
control signals, the output current of the servo amplifier is measured by a current
probe to determine the gain and phase shift of the current controller at different
frequency range. The current loop response of the servo amplifier using the tracking
current of 1.25A is shown in Fig. 6.10.
94
Fig. 6.10 Frequency response of the current loop control of the servo amplifier
In Fig. 6.10, the roll-off of the gain in the current control loop commenced at 2 KHz
and the current loop bandwidth with cut-off frequency of -3dB is 7.42 KHz. The
phase shift of the current control loop at 1 KHz is only -8.72. Hence, the servo
amplifier can perform approximately as a pure amplification device without adding
extra delay and phase shift to the closed-loop position control system of the parallel
manipulator.
6.4 Experimental Verification of the Dynamic

Model
The dynamic model of the 4-DOF parallel manipulator has been simulated using
sinusoidal force profiles given in Chapter 5. The mathematical model of the parallel
manipulator can be verified experimentally using the physical manipulator prototype.
The manipulator prototype is driven using the same force profiles given by (5.22)(5.25) and the actuator positions are measured by the linear encoders. The difference
95
between the simulation position output and the actual positions of the linear actuators
are plotted in Figures 6.11-6.14.
Fig. 6.11 Difference on the simulated and measured actuator positions for case (,)
= (89.8, 30)
= (89.8, 330)
96
Fig.6.13 Difference on the simulated and measured actuator positions for case (,)
= (125.7, 150)
= (125.7, 210)
97
The trajectories of the end-effector computed from the experimental results are
plotted as solid curves on Fig. 6.15. Figures 6.11-6.15 show that the experimental
results match the simulated results very well and the trajectories are almost
indistinguishable. Comparing Figures 6.11-6.12 with Figures 6.13-6.14, we see that
although the motion magnitude is smaller at a higher frequency due to smaller
system gain, the position tracking accuracy remains good as long as the input profile
rise time is within the specified limit of the manipulator.
Fig. 6.15 Simulated and experimental trajectories of the end-effector
6.5 Summary
A prototype of the 4-DOF parallel manipulator is constructed to realize the design of
the manipulator using the approaches developed in the previous chapters. The
parallel manipulator is driven by direct-drive linear in AC servo motors such that the
overall moving mass of the manipulator can be minimized and the energy loss in
power transmission when compared with the traditional rotary motor system can be
avoided. The structure of the motion control platform of the parallel manipulator is
described, whereby the kinematics and trajectory calculations are performed in the
98
host PC and the closed-loop control algorithm of the manipulator are implemented in
the DSP controller. The current control loop in the PWM servo amplifier is designed
and verified using frequency domain method to ensure that the tracking response of
the current controller is adequate with gain linearity and minimum phase shift. The
dynamic model of the parallel manipulator is verified by an open-loop approach
using sinusoidal force input profiles, where the actuator positions are measured by
linear encoders. From the position difference between numerical simulation and
experimental results, the dynamic model obtained in Chapter 5 can fairly accurately
describe the trajectory response of the physical prototype.
99
Chapter 7
Control Design of the Planar Parallel
Manipulator
7.1 Introduction
We will study in this chapter, the design of the closed-loop control system for the 2DOF planar parallel manipulator that drives the 4-DOF parallel manipulator. The
planar manipulator is interesting in its own right as it can be used as a superior
alternative to the XY motion stage. First, a traditional PID computed-torque control
system is designed using the model-based approach [43], [44]. The mathematical
model of the planar mechanism developed in chapter 5 is linearized using feedback
linearization method. However, the computed-torque control is inadequate to provide
the required motion performance due to the limited controller gain. To overcome the
weakness of the computed-torque control, a robust learning control method is
developed to minimize the settling time and steady-state error under repetitive pointto-point motion. A frequency-domain system identification approach is developed to
determine the dynamic model of the manipulator [48], [49], [50]. A robust control
design method [51] is employed to design a stable, fast tracking-response feedback
controller with less sensitivity to high frequency disturbance and the control
parameters are determined using genetic algorithm. A Fourier-series based iterative
learning controller [65], [66], [67] is designed and used on the feedforward path of
the controller to further improve the settling time by reducing the dynamic tracking
100
error of the manipulator. Experimental results demonstrate that the robust learning
controller has significant improvements in the position accuracy and settling time of
the end-effector when compared to the computed-torque control method.
Independent
end-point
position
measurements
are
performed
using
laser
displacement measurement system to show that the planar parallel manipulator

outperforms existing XY-stages by a significant margin.
7.2 Computed-torque Control Design

The kinematics of the planar parallel manipulator is shown in Fig. 7.1.
Fig. 7.1 Kinematics design of the 2-DOF planar parallel manipulator
The free-body diagram of the planar parallel manipulator is already shown in Fig. 6.1.
In the free-body diagram, m1 and m3 are the masses of the linear actuators and m2 and
m4 are the masses of the linkages. The inertia of link 2 and link 4 about the centre of
mass are I2 and I4 respectively. The dynamic model of the planar parallel manipulator
has already been expressed in (5.11). In the planar parallel manipulator model, the
external forces p and p are the reaction from the end-effector with mass mp and
can be expressed as
101
p = m p y1
p = m p
x1
and
and the acceleration of x1 and y1 can be expressed in the form of the accelerations of
the linear actuators obtained from d1 and d3. Hence, the model expressed in (5.11)
can be simplified as
Fp = M p ( q ) q + C p ( q, q )
(7.1)
where Fp denotes the actuating force vector, q represents the generalized coordinates
vector, Mp represents the mass matrix of the planar manipulator and Cp represents the
velocity coupling vector.
In the parallel manipulator, the inputs are the force exerted by the linear actuators,
Fp = [ F1
F3 ]
(7.2)
and the generalized joint variable vector is
q = [ d1
d3 ]
(7.3)
Because of the closed kinematics structure and also by the symmetry of the linkages,
the joint variables of the passive revolute joints can be expressed in terms of q as:
2 = 4 = cos 1 ( ( d3 d1 ) / 2l )
The parameter matrices in (7.1) are given by:
M
M p = 11
M 21
M 12
M 22
102
(7.4)
C p = [C1 C2 ]
m2 r2 m4 r4 2 + I 4 + m2 r2 2 + I 2 m p h2 c 2 m p
+
l2
4l2 2 s 2 2
2 s 2
4
M 11 = m1 + m2
M 12 =
(7.5)
m4 r4 + m2 r2 m4 r4 2 + I 4 + m2 r2 2 + I 2 m p h2 c 2 m p
2l2
4l2 2 s 2 2
2 s 2
4
M 21 = M 12
M 22 = m3 m4 +
C1 = C2
(m r
=
4 4
h1 =
+ I 4 + m2 r2 2 + I 2 ) c 2 d1 d3
8l23 s 4 2
( d
d1
2
16 l ( d 3 d1 ) / 4
2
h2 =
4
(7.7)
(7.8)
m4 r4 m4 r4 2 + I 4 + m2 r2 2 + I 2 m p h2 c 2 m p
+
+
l2
4l2 2 s 2 2
2 s 2
4
( d3 d1 )
(7.6)
3/ 2
( d
m p h1c 2
d1
2 s 2
(7.9)
(7.10)
(7.11)
4 l ( d 3 d1 ) / 4
2
( d3 d1 )
2
l2 2 ( d3 d1 ) / 4
(7.12)
The dynamic model of the planar parallel manipulator expressed in (7.1) is used for
designing the PID computed-torque controller. First, feedback linearization is applied
to convert the nonlinear manipulator model into a linear model. Suppose the
actuating force vector Fp is generated by
Fp = M p ( q ) u + C p ( q, q )
where u is a new control input. Equating (7.1) and (7.13) yields
103
(7.13)
M p ( q ) q + C p ( q, q ) = M p ( q ) u + C p ( q, q )
q = u
(7.14)
Fig. 7.2 Feedback linearization of the manipulator model
Fig. 7.2 shows a block diagram of the feedback-linearized manipulator model. Let
qd be the desired joint acceleration of the linear actuator and v be the control output
of the position feedback loop. In this case, the control input u can be expressed as
u = qd + v
A joint-based PID computed-torque controller is implemented to control the position

of the planar manipulator. The PID controller is expressed in the form of
v = K P e + K I edt + K D e
(7.15)
where e is the position error between the command position qd of the linear actuator
and the joint position q acquired from the linear encoder. Substituting (7.15) into
(7.14), the overall control law becomes
Fp = M P ( q ) qd + K P e + K I edt + K D e + C p ( q, q )
104
(7.16)
Fig. 7.3 shows the block diagram of the proposed computed-torque PID controller.
The closed-loop characteristic polynomial defined in [76] of the PID computedtorque controller described in (7.16) is
c ( s ) = s3 I + K D s 2 + K P s + K I
(7.17)
and the diagonal control gains are defined as K D = diag {K Di } , K P = diag { K Pi } ,
K I = diag { K Ii } (i = 1 and 3). The damping ratio and the natural frequency n of
the joint error i are used for designing the control gains of the manipulator as
K Pi = n 2 ,
K Di = 2n ,
and
K Ii < K Di K Pi
Fig. 7.3 Block diagram of the PID computed-torque controller
7.3 Robust Control Design

The PID computed-torque control method developed in section 7.2 is inadequate to
provide the required motion performance due to the limitation on the controller gain.
As the controller gain of the PID computed-torque is increased to reduce the dynamic
and steady-state error of the manipulator, the control bandwidth is increased and the
mechanical resonant of the manipulator can be excited by the high frequency
components of the control signal. Furthermore, with increased bandwidth, high
frequency noise is introduced into the control loop and the stability and robustness of
the control system of the planar manipulator are degraded. Hence, robust control is
essential to maintain closed-loop stability and the position accuracy of the
105
manipulator. To overcome the weakness of the PID computed-torque controller, a

linear time-invariant model is required to develop the design of the robust feedback
control system.
Second-order linear time-invariant model is commonly used to represent the lowfrequency dynamics of servo mechanisms. However, it is inadequate in the planar
manipulator model as the high-frequency dynamics is critical to the motion
performance and stability of the high speed and high precision parallel manipulator.
It is essential to include the high-frequency dynamics in the design of control system.
Hence, we will represent the planar parallel manipulator using a high-order jointspace dynamic model of the form
qi ( s )
ui ( s )
= Pi ( s ) = Li ( s ) H i ( s )
Li ( s ) =
Ki
s ( s + i )
2
2
rm
s 2 + 2 amam s + am
H i ( s ) = 2 2
2
m =1 am s + 2 rmrm s + rm
(7.18)
(7.19)
(7.20)
where Pi(s) (i = 1 and 3) is the joint-space dynamic model of the planar manipulator,
qi(s) is the position output of the translational joints, ui(s) is the controller input of the
manipulator, Li(s) is the low frequency model and Hi(s) is the high frequency model
of the manipulator, Ki is the system gain and i is the viscous friction coefficient of
the translational joints, M is the number of resonant frequency included in the high
frequency model, rm and am are the mth resonant frequency and anti-resonant
frequency of the translational joint i with the damping ratios rm and am respectively.
A low bandwidth motion trajectory is selected to perform a point-to-point motion on

the translational joints. The controller input and the translational joints position
106
feedback are acquired to determine the model parameters of the low frequency model
Li(s) expressed in (7.19) using prediction error method (PEM).
The high frequency model Hi(s) in (7.20) can be identified using the sweep-sine
method. In this method, sinusoidal signal with desired peak amplitude over a specific
frequency range is generated to excite the open-loop mechanism. The translational
joints acceleration is estimated from the encoder feedback position. The gain and
phase of the high frequency model can be determined using the acceleration output
and the input of the translational joints. The high frequency model Hi(s) in (7.20) is
then estimated using least squares estimation (LSE) method.
A joint-based robust controller is designed to control the position of the planar

parallel manipulator as shown in Fig. 7.4.
Fig. 7.4 Block diagram of the robust controller
The feedback path consists of a PID controller Gc(s), a lag compensator FL(s) and a
notch filter FN(s) which can be expressed as
Gc ( s ) =
uc ( s )
e(s)
(s K
=
FL ( s ) =
+ sK P + K I
2
s +s
KL ( s + z )
(s + p)
107
(7.21)
(7.22)
FN ( s ) =
s 2 + (h l ) s + (h l )
s 2 + (h l ) s + (h l )
(7.23)
where uc(s) is the output of the PID controller, e(s) is the position error between the
desired position qd(s) of the translational joint and the joint position q(s) acquired
from the linear encoder, is a time constant, KL is the normalized gain of the lag
compensator, p = -exp(-2fp) is the pole of the lower corner frequency fp and z = exp(-2fc) is the zero of the higher corner frequency fc, is the attenuation of the
notch filter measured from the resonant peak to the flat band of the plant gain with
high break-frequency h and low break-frequency l.
The feedforward path of the proposed robust controller can be expressed as
u f (s)
qd ( s )
= G f ( s ) = K j s3 + Ka s 2 + Kv s
(7.24)
where uf(s) is the control output of the feedforward path, Kv is the velocity
feedforward constant, Ka is the acceleration feedforward constant and Kj is the jerk
feedforward constant of the translational joint.
Using the equations (7.21) to (7.24), the overall control law of the joint-based robust
controller becomes
u ( s ) = FL ( s ) FN ( s ) ( uc ( s ) + u f ( s ) )
(7.25)
Closed-loop stability, sensitivity to disturbance, robustness and fast tracking response

are essential to the robust control design of the high-precision manipulator.
Accordingly, the gain margin (GM), the phase margin (PM), the closed-loop control
bandwidth (CLBW) and the sensitivity function (S) of the proposed robust controller
are optimized to achieve the requirements described in [80]:
108

1
GM = 20 log
Gc FL FN P ( j )
7dB
GC FL FN P( jw)=180
PM = Gc FL FN P ( j ) G F F
c L NP
( j ) =1
(7.26)
60
(7.27)
CLBW = min ( ) (Gc + FFC ) FL FN P ( j ) 10
(7.28)
1+ Gc FL FN P ( j )
max S ( jw ) CLBW = max
1
<1
1 + Gc FL FN P ( jw ) CLBW
(7.29)
In (7.28), the closed-loop bandwidth definition of the proposed control algorithm is

refined to ensure the tracking response is not affected by the system phase lag.
An objective function is defined to take into account the gain margin difference GM,
phase margin difference PM, closed loop bandwidth difference CLBW and the
maximum value of S over the CLBW, given by:
J ( K P , K I , K D , K v , K a , K j , K L , f c , f p , , l , h )
= GM + PM + CLBW + max S ( j ) CLBW
(7.30)
The control parameters of the proposed robust controller are then designed by
solving the following minimization problem using genetic algorithm:
min
J K P , K I , K D , K v , K a , K j , K L , f c , f p , , l , h
( K P , K I , K D , Kv , Ka , K j , K L , fc , f p , ,l ,h )
7.4 Robust Learning Control Design

In semiconductor packaging system, the manufacturing operations are mainly
repetitive motion between two desired positions. The reduction on the dynamic
109
tracking error of the translational joints can reduce the settling time of the endeffector at the end of motion. However, the increase of the closed-loop controller
gain is constrained by the mechanical bandwidth of the manipulator. To avoid the
use of excessive gains, a robust learning controller is developed to enhance the
dynamic tracking performance of the linear actuators of the manipulator. Fig. 7.5
shows the proposed learning controller consisting of two components: (i) the robust
feedback controller developed in section 7.3 for closed-loop stability, and (ii) a
learning controller in the feedforward path for reducing tracking position error.
Fig. 7.5 Block diagram of the robust learning controller
The operation of the proposed learning controller is described in Fig. 7.6. The motion
time for a complete point-to-point travel Tm is defined as the sum of the trajectory
time Tp and the desired settling time Tt. The delay time between two motion cycles is
defined as Td. The position error of the translational joint is acquired by the robust
feedback controller in real-time with the sampling time of Ts. At the end of a
complete motion cycle j, the position error sequence is uploaded to the offline
learning controller to determine the feedforward learning control.
110
Fig. 7.6. Functional diagram of the learning controller
The learning control design is shown in Fig. 7.7. Assume the position error sequence
is a continuous periodic signal e(t) defined over the time interval of [0,Tm). The
sequence is sampled and represented by Discrete Fourier Transform (DFT) as
N 1
e(kTs ) = Ro + ( Rn cos n kTs + I n sin n kTs )
(7.31)
n =1
Ro =
Ts
Tm
N 1
e(kT ) , R
s
k =0
2Ts
Tm
N 1
e(kT ) cos nkT

s
k =0
, In =
2Ts
Tm
N 1
e(kT ) sin n kT
s
k =0
where N is the number of samples in the interval of [0,Tm), n = 1,2,N-1, k =

0,1,2,N-1, 1/Tm denotes the base frequency of the Fourier series, =2/Tm is the
frequency in rad/s, Ts is the sampling time, and Rn and In are the coefficients of the
Fourier series.
Fig. 7.7 Block diagram of the learning controller
111
In Fig. 7.7, the Fourier coefficients of the error sequence can be obtained using FFT.
A FFT lag filter and a FFT notch filter are used to remove the high frequency and
resonant frequency components of the error such that the manipulator will not be
excited by the disturbances in the control loop. Zero phase delay is another benefit of
the FFT filters which does not introduce additional phase shift in the control loop.
The iteration law of the learning loop and the control output of the learning controller
are formulated as
j +1 (k ) = Gl j +1 (k ) + j (k )
Kd
Kd

j +1 ( k 1)
j +1 ( k )
j +1 ( k 1) + K p +
Ts
Ts
1
j +1 ( k ) =
(7.32)
(7.33)
where j (k ) and j +1 (k ) are the inverse Fourier transform of the filtered spectrum
of the error signal in the present cycle j and the next cycle (j+1), Gl is the learning
loop gain with (0 < Gl 1.0) and is the forgetting factor of the learning loop with (0
< 1.0) to adjust the convergence rate, j +1 (k ) is the control output before
regulation and j +1 (k ) is the regulated iterative learning control output by the PD
controller. The control output j +1 (k ) is stored in the learning controller and
downloaded to the real-time feedback controller which is synchronized to the
beginning of the next motion cycle (j+1). The feedforward control output in (7.24) is
modified as
u f ( k ) = ( K v + K a + K j ) qd ( k ) qd ( k 1) ( K a + K j ) qd ( k 2 ) K j qd ( k 3) + j +1 ( k )
The stability of a simple PD plus learning control algorithm has been proved in [81].
In the robust learning controller, as the learning loop output signal uf converges to the
controller signal us, the output of the PID controller Gc tends to zero which implies
the convergence of the position tracking error e. Hence, the position tracking error
can be minimized by the proposed learning controller with closed-loop stability and
robustness.
112
7.5 Experimental Results

A prototype of the planar parallel manipulator was designed and constructed as
shown in Fig. 7.8 with target applications for semiconductor packaging. A typical
BGA substrate is mounted on the end-effector to simulate the loading in actual
semiconductor packaging operations. As the parallel manipulator is intended for
high-precision alignment of semiconductor chips substrate whereas the longer range
transportation operation will be performed by other mechanisms, the required
workspace of the manipulator is small (of size 15 mm 15 mm ). The dimensions and
the mechanical characteristics of the linkages are given in Table 6.1. The actual
workspace of the mechanism is of a diamond shape (about 40 mm horizontally across
at its widest, and 20.3mm from top to bottom). In Fig.7.8, a high precision laser
displacement measurement system Keyence LB-041 with 0.91m /mV resolution is
mounted on the fixed base of the manipulator to provide an independent
measurement of the end-effector position.
BGA substrate
Laser
measurement
system
End-effector
Position
encoder
Position
encoder
Ironless linear brushless

servo motors
Fig. 7.8 Experimental setup of the planar parallel manipulator
113
A motion trajectory is designed to carry out a typical point-to-point motion of the

planar manipulator for performing the substrate alignment process. The original
location of the end-effector measured from the origin is (x0, y0) = (109.5mm,
73.246mm) and the destination is (x1, y1) = (117mm, 74.372mm). The required
trajectory time of this motion is 70ms. The end-effector trajectory is mapped into two
5th order polynomial motion profiles for the linear actuators and pre-computed using
the inverse kinematics model given in (2.11) and (2.12). The PID computed-torque
controller developed in section 7.2 and the robust controller designed in section 7.3
are downloaded into the DSP platform for real-time control with sampling frequency
1/Ts of 2 KHz. In the robust learning controller, the learning loop is implemented in
the host PC in an offline basis. The control output of the learning loop is then
downloaded to the DSP platform and synchronized with the feedback position
control loop in real-time control. The required end-point accuracy of the end-effector
to the destination point is 5m in the XY plane with a desired setting time Tt of
10ms after the end of motion.
The joint based robust controller is designed by the dynamic model in (7.18). The
low frequency models are identified by PEM identification using two low speed
motion trajectories of 10mm in 70ms for L1 and 10mm in 80ms for L3 respectively.
The measured encoder positions of the translational joints are compared with the
simulated position output using the PEM model and shown in Figs. 7.9 and 7.10.
114
Fig. 7.9 Model matching of the d1 actuator using PEM identification

The low frequency dynamic models can be expressed as
115
L1 ( s ) =
11440
and
s ( s + 11.767)
L3 ( s ) =
10797
s ( s + 22.22)
(7.34)
The high frequency models H1(s) and H3(s) for joints 1 and 3 are determined by the
swept-sine identification using a sinusoidal control signal with the frequency range
from 80Hz to 500Hz to excite the mechanism. The frequency response of the
translational joint is acquired by placing the linear motor pair at different positions.
In Figs. 7.11 and 7.12, the worst case frequency response of the translational joint
with the lowest initial gain is selected for the controller design.
Fig. 7.11 Frequency response of the planar parallel manipulator in d1 joint space
116
Fig. 7.12 Frequency response of the planar parallel manipulator in d3 joint space
From the experimental results, the high frequency dynamics can be represented by 4th
order models expressed as
H1 ( s ) =
and
)( s
0.49715 s 2 + 74.07 s + 2.59 106
H 3 (s) =
(s
+ 86.24s + 2.67 106
)( s
0.13805 s 2 88.36s + 2.584 106
(s
+ 79.18s + 2.286 106
)( s
945.9s + 1.195 107
+ 283s + 5.763 106
)( s
2841s + 4.821 107
+ 133.9s + 7.567 106
(7.35)
Substituting (7.34) and (7.35) into (7.18), the joint based models of the planar
parallel manipulator are determined to be
P1 ( s ) =
and
5687.4 s 2 + 74.07 s + 2.59 106
)( s
s ( s + 11.767 ) s + 86.24s + 2.67 10

2
945.9s + 1.195 107
)( s
+ 283s + 5.763 106
)( s 2841s + 4.82110 )
P ( s) =
s ( s + 22.22 ) ( s + 79.18s + 2.286 10 )( s + 133.9s + 7.567 10 )
1490.53 s 2 88.36s + 2.584 106
117
The parameters of the robust controller are then designed by using genetic algorithm
and are listed in Table 7.1.
d1 actuator controller
d3 actuator controller
KP
1.4429
0.48379
KI
0.0038175
0.000222
KD
12.424
6.5531
Kv
4.562
9.8422
Ka
22.232
10
DC gain
10
10
F1 (Hz)
331
302
F2 (Hz)
825
891
r (Hz)
371
438
b (Hz)
200
300
10
23
G.M. (dB)
8.64
7.54
P.M. (deg)
48.4
57.5
C.L.B.W. (Hz)
75.91
41.85
0.81
0.86
Parameter
Controller properties
Table 7.1 Parameters of the robust feedback controller designed by Genetic

Algorithm
The gain margin and phase margin of the open-loop system of joints d1 and d3 are
shown in Figs. 7.13 and 7.14 respectively. The learning loop gain Gl is chosen to be
0.8 and the forgetting factor is set to be 0.995 in order to obtain a stable learning
loop with adequate rate of convergence.
118

The sensitivity function S of the robust controller of d1and d3 actuators are shown in
Figs. 7.15 and 7.16 respectively.
119
Fig. 7.15 Sensitivity function of the robust controller of d1 actuator

Figs. 7.17 and 7.18 show the position and velocity tracking of the actuator d1
compared with the 5th order polynomial motion profiles generated in the joint space
using different type of controllers.
120

The tracking performance of the actuator d3 is shown in Figs. 7.19 and 7.20.
121

The position errors of the actuators d1 and d3 as measured by the linear encoders in
the joint space are shown in Figs. 7.21 and 7.22.
122

The convergence the maximum dynamic position error of d1 and d3 actuators under
robust learning control and the required number of iteration are shown in Fig. 7.23.
123
Fig. 7.23 Convergence of maximum dynamic error of d1 and d3 actuators using

From Fig. 7.23, the number of iterations required for the convergence of the
maximum dynamic position error of the linear actuators d1 and d3 is 12 cycles.
The desired and estimated trajectory of the end-effector are determined using the
forward kinematics model (2.5) and (2.6) derived in chapter 2 and plotted in Fig.
7.24. The estimated position errors of the end-effector in the X and Y directions are
shown in Figs. 7.25 and 7.26.
124
Fig. 7.24 Position tracking of the end-effector in the Cartesian space determined by
the forward kinematics model
Fig. 7.25 Position error of the end-effector in X-direction of the Cartesian space
125
Fig. 7.26 Position error of the end-effector in Y-direction of the Cartesian space
Table 7.2 demonstrates the improvement on the motion performance of the planar
parallel manipulator using the robust learning controllers in terms of maximum
absolute dynamic position error, settling time and steady-state error of the linear
actuators.
Computed-
Robust
Robust
torque control
control
learning
control
Max. absolute dynamic error of d1
136.4
13.4
4.4
84.4
14.0
9.4
Settling time of d1 (ms)
36.5
2.0
9.5
Steady-state error of d1 (m)
-1.5
-0.2
-1.0
-1.0
(m)
Max. absolute dynamic error of d3
(m)
Table 7.2 Comparison of the maximum absolute dynamic error, settling time and
steady-state error of the linear actuators using different controllers
126
With the benefits of the robust control design method and the Fourier based learning
mechanism, the dynamic position error and the settling time of the linear actuators
are reduced significantly compared to the traditional computed-torque control.
Whether the end-effector achieves the desired end-point accuracy within the desired
settling time is also verified independently using a laser displacement measurement
system. The X and Y positions of the end-effector is measured over an interval
covering the end of motion profile of the linear actuator (dotted vertical line
corresponds to Tp =70ms) and the time when the position accuracy falls within 5m.
Fig. 7.27 shows the settling response and the stead-state error of the end-effector
using the traditional computed-torque control method.
(b) Y-direction
(a) X-direction
Fig. 7.27 Position response of the end-effector (grey) and the driving current (black)
over the settling interval using Computed-torque Control
The robust controller is then used to compare settling responses of the end-effector
with the computed-torque controller and shown in Fig. 7.28.
127
(b) Y-direction
(a) X-direction
over the settling interval using Robust Control
The benefits of the robust learning controller on the further improvement of the
settling time of the end-effector after the end of motion is shown in Fig. 7.29.
(a) X-direction
(b) Y-direction
over the settling interval using Robust Learning Control
The measurement results of the planar manipulator are compared with traditional XY
motion stages mounted with the same BGA substrate. A significant reduction in the
settling time of the end-effector is obtained as compared with the traditional XY
motion stage given in Table 7.3.
128
XY Motion
Planar
Planar
Planar
Stage
Manipulator
Manipulator
Manipulator
(Robust
(Computed
(Robust
(Robust
Control)
-torque
control)
learning
control)
Settling time in X-
control)
28
48
16.0
8.8
22
60
6.0
1.82
1.37
-0.91
8.0
2.73
1.09
0.91
direction (ms)
Settling time in Ydirection (ms)
End-point accuracy in
X-direction (m)
Y-direction (m)
Table 7.3 Comparisons of the settling time and end-point accuracy (as determined
by laser displacement measurement system) of the planar manipulator using different
controllers and the XY motion stage
From the results, the settling time of the planar manipulator, in comparison with the
XY-stage under the same kind of robust feedback control scheme, has been reduced
form 28ms to 16ms in X-direction and 22ms to 0ms in Y-direction, and the steadystate error has been reduced from 6m to 1.37m in X-direction and 8m to 1.09m
in Y-direction. The performance of the planar manipulator is further enhanced using
the proposed robust learning control scheme. The settling time is reduced from 16ms
to 8.8ms in X-direction and the steady-state error is reduced from 1.37m to -0.91m
in X-direction and 1.09m to 0.91m in Y-direction. These experimental results
using independent laser measurement equipment demonstrated that the proposed
planar manipulator under the robust learning control scheme provides a superior
alternative to the XY motion stages for high precision positioning.
Table 7.4 shows further advantages of the proposed manipulator over the traditional
XY motion stage in terms of the total moving mass, the required peak current and the
129
required motor power. Other benefits include a reduction in the size of the
mechanism and savings in the actuator driving energy.
XY Motion
Planar
Stage
Manipulator
Moving mass in X-direction (kg)
8.2
0.97
Moving mass in Y-direction (kg)
3.0
0.97
Required actuator peak current (A)
12
Required actuator power (W)
100
20
Table 7.4 Comparison of the moving mass and the actuator power of the planar
manipulator and the XY stage
Table 7.5 shows a comparison of mechanism size, travelling range of the actuators
and the size of workspace envelope of the traditional XY stage and the planar
manipulator.
XY Motion
Planar
Stage
Manipulator
Size of the mechanism (mm x mm)
460 x 460
210 x 270
Travelling range of actuators (mm)
40
50
40 x 40
40 x 21
Workspace envelope (mm x mm)
Table 7.5 Comparison of the mechanism size, travelling range of the actuators and
the workspace envelope volume of the planar manipulator and the XY stage
7.6 Summary
A planar parallel manipulator has been developed for point-to-point motion in
semiconductor packaging applications using the basic triangular mechanism module
extracted from the 4-DOF parallel manipulator. Traditional PID computed-torque
controller is implemented based on the feedback linearized dynamic model to
130
provide a feedback control system but the tracking response and the steady-state
position accuracy of the planar manipulator under such a control scheme is found to
be inadequate. A robust controller structure is therefore proposed that allows higher
control gain to be used to improve the tracking performance of the manipulator in
dynamic motion without exciting the high-frequency resonant modes of the
mechanism. On the basis of the robust controller, a robust learning controller is
developed to further reduce the dynamic tracking error of the manipulator by means
of off-line iterative adaptation against the position error pattern. The performance as
measured by the high-precision laser displacement system provides an independent
and convincing verification of improvements in the accuracy of the end-effector,
which also shows the effectiveness of the proposed control design method.
Furthermore, the size of the manipulator and the driving energy of the actuators for
the planar parallel manipulator are less demanding than those required for the
traditional XY motion stage. From the experimental results, the planar parallel
manipulator provides a superior alternative to the XY motion stage for applications
in semiconductor packaging systems.
131
Chapter 8
Control Design of the 4-DOF Parallel
Manipulator
8.1 Introduction
The control systems developed in Chapter 7 for the planar manipulator will be
extended in this chapter to control the end-point positioning of the 4-DOF parallel
manipulator. The PID computed-torque control system is designed using the modelbased approach and the mathematical model of the 4-DOF parallel manipulator is
linearized using feedback linearization method. Similar to the case of the planar
manipulator, it is found that the tracking performance and settling time of the
manipulator fall short of the required level because of limitations on controller gain.
The robust learning control method is again employed to minimize the settling time
and steady-state error of the manipulator using a robust feedback controller and a
Fourier-series based learning feedforward controller for repetitive point-to-point
motion. Experimental results demonstrate that the robust learning controller has
significant improvements in the position accuracy and settling time of the endeffector when compared to the computed-torque control method. Independent endpoint position measurements are performed using laser displacement measurement
system to show that the 4-DOF parallel manipulator outperforms existing positioning
mechanisms with individual XY, Z and stages by a significant margin.
132
8.2 Computed-torque Control Design

The dynamic model of the 4-DOF parallel manipulator expressed in (6.11) is used
for designing the computed-torque controller. Feedback linearization is applied to
convert the nonlinear manipulator model into a linear model. Suppose the driving
force vector Fp is generated by
Fp = M p ( q ) u + C p ( q, q ) + D p ( q ) Fdv
(8.1)
where u is a new control input. Equating (8.1) and (6.11) yields
M p ( q ) q + C p ( q, q ) + Dp ( q ) Fdv = M p ( q ) u + C p ( q, q ) + D p ( q ) Fdv
(8.2)
That is,
q = u
Fig. 8.1 Feedback linearization of the 4-DOF parallel manipulator model
Fig. 8.1 shows the block diagram of the linearized model of the 4-DOF parallel
manipulator. Let qd be the desired acceleration of the linear actuator and v be the
control output of the position control loop. Hence, the control input u can be
expressed as
u = qd + v
133
The joint-based PID computed-torque controller expressed in (7.15) is employed to

control the parallel manipulator. Substituting (7.15) into (8.2), the overall control law
becomes
Fp = M p ( q ) qd + K P e + K I edt + K D e + C p ( q, q ) + Dp ( q ) Fdv
(8.3)
where e is the position error between the desired position qd of the linear actuator
and the joint position q acquired from the linear encoder feedback. The block
diagram of the PID computed-torque controller of the 4-DOF parallel manipulator is
shown in Fig. 8.2.
Fig. 8.2 Block diagram of the PID computed-torque controller of the 4-DOF parallel
manipulator
The closed-loop characteristics polynomial of the PID computed-torque controller in

(7.17) can be used for designing the control gains of the 4-DOF parallel manipulator.
In this case, the diagonal control gains are defined as K D = diag {K Di } ,
K P = diag { K Pi } , K I = diag { K Ii } (i = 1,3,5 and 7). The control gains of the 4-DOF
parallel manipulator can be designed using the damping ratio and the natural
frequency n of the joint error i as
K Pi = n 2 ,
K Di = 2n ,
134
and
K Ii < K Di K Pi
8.3 Robust Control Design

The robust feedback controller of the planar parallel manipulator described in section
7.3 is employed to improve the stability and controller gain limit of the 4-DOF
parallel manipulator. The high order joint-space dynamic model of the manipulator is
developed using (7.18) to (7.20). In (7.18), Pi(s) (i = 1,3,5 and 7) is the joint-space
dynamic model of the 4-DOF parallel manipulator relating the position output of the
translational joints qi(s) to the controller input ui(s). The low frequency model Li(s) is
determined from a low bandwidth point-to-point motion using prediction error
method. The high frequency model Hi(s) is identified from the swept-sine method to
acquire the translational joints acceleration which is excited by the sinusoidal signal
input and estimated by a least squares estimation method.
The control law of the joint-based robust controller expressed in (7.25) can be
applied to control the 4-DOF parallel manipulator. The gain margin, phase margin,
closed-loop bandwidth and the sensitivity function described in (7.26) to (7.29) are
essential to the robust controller design to ensure the stability, sensitivity, robustness
and tracking response of the manipulator. The genetic algorithm is used to design the
control parameters of the robust controller by means of the minimization of the
combined objective function (7.30) defined in terms of GM, PM, CLBW and the
maximum value of S over the CLBW.
8.4 Robust Learning Control Design

The robust learning controller developed in section 7.4 for the planar manipulator is
modified to enhance the dynamic tracking performance of the linear actuator of the
4-DOF parallel manipulator. The robust feedback controller developed in section 8.3
is used to maintain the closed-loop stability and the learning feedforward controller
is developed in this section for reducing tracking position error.
The position error sequence of the translational joints di (i = 1,3,5 and 7) are acquired
by the feedback controller and uploaded to the offline learning controller to
135
determine the learning feedforward control. The position error sequence is sampled
and represented by the DFT as (7.31) and the Fourier coefficients of the error
sequence are determined using FFT. The FFT lag filter and the FFT notch filter are
used to eliminate the high frequency and resonant frequency disturbance to the
manipulator. The iteration law and the output of the learning controller are expressed
in (7.32) and (7.33) respectively. The learning controller output is downloaded to the
real-time feedback controller which is synchronized to the beginning of the next
motion cycle.
8.5 Experimental Results

The prototype of the 4-DOF parallel manipulator with a typical BGA substrate
mounted on the end-effector to simulate the actual loading in semiconductor
packaging operations is shown in Fig. 8.3. The dimensions and the mechanical
characteristics of the linkages have readily been given in Table 6.1. The actual
workspace of the 4-DOF parallel manipulator is about 35mm in x-axis, 35mm in yaxis and 6mm in z-axis. In Fig. 8.3, the laser displacement system is mounted on the
same fixed base of the manipulator to provide an independent measurement of the
end-effector position.
BGA
substrate
End-effector
Laser
measurement
system
Fig. 8.3 Experimental setup of the 4-DOF parallel manipulator
136
A motion trajectory is designed to carry out a point-to-point motion of the 4-DOF

parallel manipulator for performing the substrate alignment process. The original
location of the end-effector measured from the origin is (x0, y0, z0, 0) = (91.4mm,
91.4mm, 123.7mm, -0.7854rad) and the destination is (x1, y1, z1, 1) = (95.1mm,
96.9mm, 122.3mm, -0.8rad) with the target trajectory time of 70ms. The end-effector
trajectory is mapped into four 5th order polynomial motion profiles for the linear
actuators using the inverse kinematics model given in (2.11) to (2.14) and (2.23) to
(2.26) and pre-computed in the PC. The PID computed-torque controller developed
in section 8.2 and the robust controller designed in section 8.3 are downloaded into
the DSP platform for real-time control with sampling frequency of 2 KHz. The
learning feedforward controller is implemented in the host PC in an offline basis.
The control output of the learning controller is downloaded to the DSP platform and
synchronized with the feedback position control loop in real-time control. The
required end-point accuracy of the end-effector to the destination point is 5m in
the XY plane, 5m in the vertical z-axis and 0.2mrad in the -axis with a desired
setting time of 10ms after the end of motion.
The joint based robust controller is designed by the dynamic model in (7.18). The
low frequency models are identified by means of PEM identification using two low
speed motion trajectories of 5mm in 70ms for L1 and L3 and 4mm in 70ms for L5 and
L7. The measured encoder positions of the translational joints are compared with the
simulated position outputs as shown in Figs. 8.4 to 8.7.
137
138

The low frequency dynamic models can be expressed as
L1 ( s ) =
7375
,
s ( s + 26.38 )
L3 ( s ) =
139
8792.2
,
s ( s + 21.03)
L5 ( s ) =
7785
s ( s + 15.21)
and
L7 ( s ) =
7909
s ( s + 41.15)
(8.4)
The high frequency models H1(s), H3(s), H5(s) and H7(s) are determined by the
swept-sine identification using a sinusoidal control signal with the frequency range
from 80Hz to 500Hz to excite the mechanism. The frequency responses of the
translational joints are acquired by placing the linear motor pair at different positions.
In Figs. 8.8 to 8.11, the worst case frequency response of the translational joint with
the lowest initial gain is selected for the controller design.
Fig. 8.8 Frequency response of the 4-DOF parallel manipulator in d1 joint space
140
141
From the experimental results, the high frequency dynamics can be represented by 4th
order models expressed as
H1 ( s ) =
H3 ( s ) =
H5 ( s ) =
H7 ( s) =
0.22 ( s 2 27110 s 6.5037 107 )( s 2 + 142.4 s + 5.629 106 )
(s
+ 28.74 s + 5.706 106 )( s 2 + 839.5s + 1.421 107 )
1.0 ( s 2 6399 s + 1.574 107 )( s 2 + 963.2 s + 1.013 106 )
(s
+ 78.95s + 8.27 106 )( s 2 + 706.4 s + 1.723 107 )
0.32 ( s 2 + 186.2 s + 4.178 106 )( s 2 4790 s + 2.768 107 )
(s
+ 164.5s + 4.728 106 )( s 2 + 410.7 s + 7.755 106 )
0.82 ( s 2 + 2215s + 5.395 106 )( s 2 3303s + 9.11 106 )
(s
+ 279 s + 5.185 106 )( s 2 + 1145s + 9.016 106 )
, and
(8.5)
Substituting (8.4) and (8.5) into (7.18), the joint based model of the 4-DOF parallel
manipulator are determined as
142
P1 ( s ) =
1622.5 ( s 2 27110 s 6.5037 107 )( s 2 + 142.4 s + 5.629 106 )
s ( s + 26.38 ) ( s 2 + 28.74 s + 5.706 106 )( s 2 + 839.5s + 1.421 107 )
P3 ( s ) =
P5 ( s ) =
8792.2 ( s 2 6399 s + 1.574 107 )( s 2 + 963.2 s + 1.013 106 )
s ( s + 21.03) ( s 2 + 78.95s + 8.27 106 )( s 2 + 706.4 s + 1.723 107 )
2491.2 ( s 2 + 186.2 s + 4.178 106 )( s 2 4790 s + 2.768 107 )
s ( s + 15.21) ( s 2 + 164.5s + 4.728 106 )( s 2 + 410.7 s + 7.755 106 )
P7 ( s ) =
6509 ( s 2 + 2215s + 5.395 106 )( s 2 3303s + 9.11 106 )
s ( s + 41.15 ) ( s 2 + 279 s + 5.185 106 )( s 2 + 1145s + 9.016 106 )
, and
The parameters of the robust controller of the 4-DOF parallel manipulator are then
designed by using genetic algorithm and are listed in Table 8.1.
Parameters
d1 controller
d3 controller
d5 controller
d7 controller
KP
0.7987
0.7229
0.9122
0.8438
KI
0.003
0.001
0.0005
0.01
KD
15.1038
12.8725
13.1532
15.4657
Kv
4.8751
5.5518
2.8185
9.6264
Ka
60.411
32.7569
31.04
49.4113
DC gain
10
10
10
10
F1 (Hz)
305
316
320
305
F2 (Hz)
533
716
690
617
r (Hz)
380
446
355
370
b (Hz)
400
300
400
300
15
25
15
30
G.M. (dB)
8.01
8.02
8.69
8.55
P.M. (deg)
50
50
49.3
60
Controller properties
143
C.L.B.W. (Hz)
69.6
73.4
59.3
56.1
0.98
0.85
0.94
0.7
Table 8.1 Parameters of the robust feedback controller of the 4-DOF parallel
manipulator designed by Genetic Algorithm
The gain margin and phase margin of the open-loop system of joints d1, d3, d5 and d7
are shown in Figs. 8.12 to 8.15 respectively. The learning loop gain Gl is chosen to
be 0.8 and the forgetting factor is set to be 0.995 in order to obtain a stable learning
loop with adequate rate of convergence.
144
145
Fig. 8.15 Frequency response of the open-loop system of d7 actuator

The sensitivity function S of the robust controller of d1, d3, d5 and d7 actuators are
shown in Figs. 8.16 to 8.19.
146
147
Fig.8.19 Sensitivity function of the robust controller of d7 actuator

Figs. 8.20 and 8.21 show the position and velocity tracking of the actuator d1
compared with the 5th order polynomial motion profiles using different type of
controllers.
148

149

The position and velocity tracking of the actuator d5 using different type of
controllers is shown in Figs. 8.24 and 8.25.
150

151

The position errors of the actuators d1, d3, d5 and d7 as measured by the linear
encoders in the joint space are shown in Figs. 8.28 to 8.31.
152
153

The desired and estimated trajectory of the end-effector are determined using the
forward kinematics model in (2.5) to (2.9) and (2.16) to (2.19) derived in Chapter 2
and plotted in Fig. 8.32. The time-based command trajectory of the end-effector in
the X, Y and Z directions are plotted in Fig. 8.33 and the direction is plotted in 8.34.
The estimated position errors of the end-effector in the X, Y, Z and directions are
shown in Figs. 8.35-8.38.
154
Fig. 8.32 Position tracking of the end-effector in the Cartesian space determined by
the forward kinematics model
Fig. 8.33 Time-based command trajectory of the X, Y and Z directions of the endeffector in the Cartesian space
155
Fig. 8.34 Time-based command trajectory of the direction of the end-effector in the
Cartesian space
Fig. 8.35 Position error of the end-effector in X-direction of the Cartesian space
156
Fig. 8.36 Position error of the end-effector in Y-direction of the Cartesian space
Fig. 8.37 Position error of the end-effector in Z-direction of the Cartesian space
157
Fig. 8.38 Position error of the end-effector in -direction of the Cartesian space
The position error convergence of d1, d3, d5 and d7 actuators which expressed in
maximum dynamic position error using robust learning control and the required
number of iteration are shown in Fig. 8.39.
Fig. 8.39 Convergence of maximum dynamic error of d1, d3, d5 and d7 actuators
using Robust Learning Control
158
From Fig. 8.39, the number of iteration required for the convergence of the
maximum dynamic position error of the linear actuators d1, d3, d5 and d7 is 18 cycles.
Table 8.2 demonstrates the improvements in the motion performance of the 4-DOF
parallel manipulator using the robust learning controllers in terms of maximum
absolute dynamic position error, settling time and steady-state error of the linear
actuators.
Computed-
Robust
Robust
torque
control
learning
control
control
Max. absolute dynamic error of d1 (m)
140.2
22.4
5.0
82.2
9.2
4.4
154.0
22.2
7.0
140.2
47.0
5.6
30.0
5.0
30.5
3.0
34.0
8.0
34.0
3.0
-1.6
-2.4
-0.6
-1.0
0.2
0.8
-0.4
0.4
-1.6
-0.6
-0.6
Table 8.2 Analysis of the maximum absolute dynamic error, settling time and steadystate error of the linear actuators using different controllers
It can be seen from Table 8.2 that, with the benefits of the robust learning control
design, the dynamic position error and the settling time of the linear actuators are
reduced significantly compared with traditional computed-torque control.
The end-point accuracy of the end-effector is verified independently using a laser

displacement measurement system. The X, Y, Z and positions of the end-effector is
159
measured over an interval covering the end of motion profile of the linear actuator
(dotted vertical line corresponds to Tp =70ms) and the time when the position
accuracy falls within 5m. Fig. 8.40 shows the settling response and the steadystate error of the end-effector using the traditional computed-torque control method.
(a) X-direction
(b) Y-direction
(c) Z-direction
(d) -direction
over the settling interval using Computed-torque Control
The robust controller is then used to compare settling responses of the end-effector
with the computed-torque controller, as shown in Fig. 8.41.
160
(a) X-direction
(b) Y-direction
(c) Z-direction
(d) -direction
over the settling interval using Robust Control
The benefits of the robust learning controller for further improving the settling time
of the end-effector after the end of motion are shown in Fig. 8.42.
(a) X-direction
(b) Y-direction
161
(d) -direction
(c) Z-direction
over the settling interval using Robust Learning Control
The measurement results of the 4-DOF parallel manipulator are compared with
traditional motion mechanisms with individual XY, Z and motion stages. A
significant reduction in the settling time of the end-effector is obtained as shown in
Table 8.3.
Individual XY,
4-DOF
4-DOF
4-DOF
Z and
Parallel
Parallel
Parallel
Motion Stage
Manipulator
Manipulator
Manipulator
(Robust
(Computed
(Robust
(Robust
Control)
-torque
control)
learning
control)
Settling time in X-
control)
28
51.2
27.2
22
49.6
36.8
18
36.8
7.2
23
54.4
24
4.8
6.0
-0.91
direction (ms)
Settling time in Ydirection (ms)
Settling time in Zdirection (ms)
Settling time in direction (ms)
162
X-direction (m)
8.0
0.91
0.91
0.91
5.5
-0.91
-0.91
4.8
-1.82
-1.82
1.82
Y-direction (m)
Z-direction (m)
-direction (m)
Table 8.3 Analysis of the Settling Time and End-point Accuracy of the 4-DOF
Parallel Manipulator Using Different Controllers and the individual XY, Z and
Motion Stage
From the results, the settling time of the 4-DOF parallel manipulator, in comparison
with the individual XY, Z and stages under the same kind of robust feedback
control scheme, has been reduced form 28ms to 27.2ms in X-direction and 18ms to
7.2ms in Z-direction, and the steady-state error has been reduced from 6m to 0m
in X-direction and 8m to 0.91m in Y-direction, and 5.5m to 0m in Z-direction
and 4.8m to -1.82m in -direction. The performance of the 4-DOF parallel
manipulator is further enhanced using the proposed robust learning control scheme.
The settling time is reduced from 27.2ms to 0ms in X-direction, 36.8ms to 0 ms in Ydirection, 7.2ms to 0ms in Z-direction and 24ms to 4.8ms in -direction. These
experimental results using independent laser measurement equipment demonstrated
that the 4-DOF parallel manipulator under the robust learning control scheme
provides a superior alternative to the individual XY, Z and motion stages for high
precision positioning.
8.6 Summary
The robust control and robust learning control systems developed for the planar
parallel manipulators have been extended and implemented to control the 4-DOF
parallel manipulator with point-to-point repetitive motion for semiconductor
packaging applications. The experimental results show that traditional PID
computed-torque controller implemented using the linearized dynamic model does
163
not provide adequate performance in tracking response and the steady-state position
accuracy for the 4-DOF parallel manipulator. The robust controller structure allows
higher control gain to be used to improve the dynamic tracking performance of the
manipulator without exciting the high-frequency resonant modes of the mechanism.
The robust learning controller can be used to further reduce the dynamic tracking
error and the settling time of the manipulator by means of off-line iterative
adaptation against the position error pattern. The performance as measured by the
high-precision laser displacement system provides an independent and convincing
verification of the improvement in the end-effector positioning accuracy. This shows
the effectiveness of the proposed control design method for the 4-DOF manipulator.
From the experimental results, the 4-DOF parallel manipulator provides a superior
alternative to the individual XY, Z and motion stages for applications in
semiconductor packaging systems.
164
Chapter 9
Conclusion and Future Research Work
9.1 Conclusion
In this thesis, a planar 2-DOF parallel manipulator and a spatial 4-DOF parallel
manipulator have been developed for high precision positioning mechanisms with
motion resolution in the unit of micrometer. The simplicity on kinematics design,
light moving mass and high structural stiffness of the parallel manipulators have
overcome the motion capability limitations of the XY motion stages on travel time
for point-to-point motion, settling time at the end of motion trajectory, end-point
accuracy and driving power demand of the actuators. Regarding structural stiffness,
the resonant frequency of the planar manipulator is found to be 410Hz and the
resonant frequency of the 4-DOF parallel manipulator is 380Hz. Compared to that of
the XY motion stages with resonant frequency of 220Hz, the resonant frequencies of
the proposed planar manipulator and the 4-DOF parallel manipulator are higher by a
factor of 1.86 and 1.73 times respectively. The proposed parallel manipulators have
significant improvement on kinematics design and end-point positioning accuracy
compared to existing parallel manipulators. Most parallel manipulators, such as
Stewart platform and H4 manipulator, contain ball joints and universal joints which
can only be manufactured to a tolerance of 0.1mm, while the tolerance for revolute
joint and translational joint can be constrained to 1m. Since our proposed parallel
manipulators use only single degree-of-freedom mechanical joints, the kinematics
165
error due to the precision constraints of multiple-DOF joints has largely been
eliminated, enabling micrometer accuracy to be achieved.
The volume of the workspace envelope of the 4-DOF parallel manipulator as

determined from the simulation result is (x, y, z) = (30mm, 30mm, 5mm) for a
manipulator size of (x, y, z) = (210mm, 210mm, 80mm). Compared to the Stewart
platform in [82] with the workspace volume of (x, y, z) = (13.7mm, 11.4mm,
13.4mm) with similar manipulator size, the workspace volume of our proposed
parallel manipulator is larger a factor of 2.2 times.
By kinematics optimization, we have managed to suppress the kinematics error

magnification to such a degree that the magnification factor is reduced to close to
one. That means that the end-effector kinematics error is made to match the error
tolerence of 1m allowed in the manufacturing of the linkages. The length of the
mechanical linkages and the dimensions of the rotational and prismatic joints of the
parallel manipulator will be increased if the maximum travelling range of the endeffector and the manipulator size are increased. It is essential to maintain the joint
accuracy and the machining accuacy of the linkages using the machining standard
[83] to minimize the kinematics error magnification. In the dynamic simulation, the
magnitude of the dynamic rotational angle of the small platfroms of the 4-DOF
parallel manipulator are within 0.03 deg and become negligible after the end of
motion. The simulation results demonstrated that the belt elasticity has negligible
effect to the end-point positioning accuracy at steady-state condition.
Compared to the performance of the traditional XY motion stages, the settling time
of the end-effector has been reduced by half to less than 10ms with the end-point
accuracy reduced to close to one micrometer. This means than the total motion time
required to complete a point-to-point motion and achieve the end-point position is
significantly reduced. The driving power of each actuator has also been reduced by a
factor of five to only 20W, which confirms that the driving energy is less demanding
because of the advantage of the load distribution in the parallel manipulator. The
loading due to the moving mass of the parallel manipulator is evenly distributed to
the linear actuators such that each actuator carries less than 1kg-loading of the total
166
moving mass of the manipulator. To conclude, the proposed parallel manipulator

design has achieved the required improvements in accuracy, speed and energy
efficiency and. provides a superior alternative to existing technology for positioning
mechanisms with micrometer accuracy.
9.2 Future Research Work

In this project, the design, modelling and control of high precision parallel
manipulators have been studied. Several areas can be further investigated.
The end-effector of the 2-DOF planar parallel manipulator with translational motion
on the XY plane can be modified by installing a miniature rotary servo motor to
provide an extra DOF of direct-drive rotational motion against the Z-axis. The
stiffness and rigidity of the planar parallel manipulator can be maintained as the mass
of the direct-drive rotary motor is small compared to the moving mass on the XY
plane. With the benefits on reducing the number of passive joints, linkages and the
limited distance between the end-effector and the encoder feedback, the positioning
accuracy and the workspace area of the modified planar parallel manipulator is
enhanced as compared with the traditional 3-DOF planar manipulators developed by
Gosselin and Angeles [84] and Mohammadi et al. [85].
A robust learning control method has been developed and used on the control of the
planar parallel manipulator and the 4-DOF parallel manipulator. The controller
design approach can be further developed to a novel hybrid position/force control
system by installing a force sensor at the end-effector and using the force feedback to
enhance the capability of the parallel manipulator to deal with the pick-and-place
motion with force requirements. The control loop will switch from position control to
force control mode when the end-effector achieves the desired contact position. The
position controller can be designed using an H approach with learning feedforward
loop to attenuate the external disturbance caused by the contact force of the endeffector so as to maintain the stability and robustness of the end-point position of the
manipulator.
167
List of publications
Following is the list of publications associated with this thesis:
1. J.W.F. Cheung, Y.S. Hung and G.P. Widdowson, Design and analysis of a
novel 4-DOF parallel manipulator for semiconductor applications, The 8th
Mechatronics Forum International Conference, pp.1358-1366, 2002.

2. J.W.F. Cheung and Y.S. Hung, Kinematics optimization for high positioning
accuracy of a 4-DOF parallel manipulator for semiconductor applications,
IEEE/ASME International Conference on Advanced Intelligent Mechatronics, pp.

1256-1261, 2003.
3. J.W.F. Cheung and Y.S. Hung, Modelling and control of a 2-DOF planar
parallel manipulator for semiconductor packaging systems, IEEE/ASME
International Conference on Advanced Intelligent Mechatronics, pp. 717-722,

2005.
4. J.W.F. Cheung and Y.S. Hung, Robust learning control of a high precision
planar parallel manipulator, submitted to Mechatronics, Elsevier, 2006.
5. J.W.F. Cheung and Y.S. Hung, Robust control of a high precision planar
parallel manipulator for semiconductor applications, accepted by 2nd Asia
International Symposium on Mechatronics, Hong Kong, 2006.
Awards
Following are the awards associated with this thesis:
1. Biographical publication in the 7th edition (2003-04) of Marquis Whos Who in
Science and Engineering for the contribution of High Precision Parallel
Manipulator Control System, Marquis Whos Who, Wilmette, Ill., USA.

2. Biographical publication in the 8th edition (2004-05) of Marquis Whos Who in
Science and Engineering for the contribution of High Precision Parallel
Manipulator Control System, Marquis Whos Who, Wilmette, Ill., USA.
168
3. Award of the International Scientist of the Year 2004 of International

Biographical Centre for the contribution of Robotics, Control and Mechatronics,
Cambridge, U.K.
169
References
[1] J.P. Merlet, Parallel Robots. Kluwer Academic Publishers, 2000.
[2] V.E. Gough, Contribution to discussion of papers on research in automotive

stability, control and tyre performance, in Proc. Auto Div., Inst. Mechanical
Engineers, 1956-57.
[3] D. Stewart, A platform with 6 degrees of freedom, in Proc. of the Institution of
Mechanical Engineers, 180 (part1, 15), pp.371-386, 1965.
[4] Giddings and Lewis, Giddings and Lewis machine tools, Fond du Lac, WI, USA,
1995.
[5] C. Gosselin and J. Hamel, The Agile Eye: A high performance three degree-offreedom camera-orienting device, in Proc. of IEEE Int. Conf. on Robotics and
Automation, pp.781-786, 1994.
[6] R. Clavel, DELTA4, a fast robot with parallel geometry, in Proc. of
International Symposium on Industrial Robots, pp.92-100, 1988.
[7] F. Pierrot and O. Company, H4: a new family of 4-dof parallel robots, in Proc.
IEEE Int. Conf. on Advanced Intelligent Mechatronics, pp. 508-514, 1999.
[8] S. Krut, M. Benoit, H. Ota and F. Pierrot, I4: A new parallel mechanism for
Scara motions, in Proc. IEEE Int. Conf. on Robotics and Automation, pp.18751880, 2003.
170
[9] L. Rolland, "The Manta and the Kanuk: Novel 4-DOF Parallel Mechanisms for
Industrial Handling," in Proc. of ASME Dynamic Systems and Control Division,
IMECE'99 Conference, Vol. 67, pp. 831-844, 1999.
[10] G. Pritschow and K. Wurst, Systematic design of Hexapods and other parallel
link systems, Annals of the CIRP 46, pp. 291-295, 1997.
[11] L.W. Tsai, Systematic Enumeration of parallel manipulators, Technical
report TR98-33, Institute for System Research, The University of Maryland,

U.S.A.
[12] J.M. McCarthy, Geometric design of linkages. New York: Springer, 2000.
[13] T. Arai, K. Cleary, T. Nakamura, H. Adachi and K. Homma, Design, analysis

and construction of a prototype parallel link manipulator, in Proc. IEEE Int.
Workshop on Intelligent Robots and Systems, pp.205-212, 1990.
[14] Y. Koseki, T. Arai, K. Sugimoto, T. Takatuji and M. Goto, Design and

accuracy evaluation of high-speed and high precision parallel mechanism, in
Proc. IEEE Int. Conf. on Robotics and Automation, pp.1340-1345, 1998.
[15] A.K. Dash, I.M. Chen, S.H. Yeo and G. Yang, Task-based configuration
design for 3-legged modular parallel robots using simplex methods, in Proc.
IEEE. Symp. on Computational Intelligence in Robotics and Automation,

pp.998-1003, 2003.
[16] W. Kuhlbusch, W. Moritz, J. Luckel, S. Toepper and F. Scharfeld, Triplanar

a new process-machine type developed by means of the mechatronics design,
in Proc. IEEE/ASME Int. Conf. on Advanced Intelligent Mechatronics, pp.514519, 1999.
[17] C.
Gosselin,
Determination
of
the
workspace
of
6-DOF
parallel
manipulators, Journal of Mechanical Design, Vol. 112, pp.331-336, 1990.
171
[18] J.P. Merlet, C. Gosselin and N. Mouly, Workspaces of planar parallel

manipulators, Mechanism and Machine Theory, Vol.33, No.1, pp.7-20, 1998.
[19] D.I. Kim, W.K. Chung and Y. Youm, Geometrical approach for the
workspace of 6-DOF parallel manipulators, in Proc. IEEE Int. Conf. on
Robotics and Automation, pp.2986-2991, 1997.
[20] I.A. Bonev and J. Ryu, Workspace analysis of 6-PRRS parallel manipulators
based on the vertex space concept, in Proc. ASME Design Engineering
Technical Conferences, pp.1-8, 1999.
[21] C. Gosselin, E. Lavoie and P. Toutant, An efficient algorithm for the graphical
representation of the three-dimensional workspace of parallel manipulators,
22nd Biannual Mechanisms Conference, pp.323-328, 1992.
[22] L.T. Wang and J. Hsieh, Extreme reaches and reachable workspace analysis of
general parallel robotic manipulators, Journal of Robotic Systems, Vol.15,
No.3, pp.145-159, 1998.
[23] F. Bulca, J. Angeles and P.J. Zsombor-Murray, On the workspace

determination of spherical serial and platform mechanisms, Mechanism and
Machine Theory, Vol.34, pp.497-512, 1999.
[24] J.A. Carretero, M. Nahon and R.P. Podhorodeski, Workspace analysis of a 3DOF parallel mechanism, in Proc. IEEE/RSJ Int. Conf. on Intelligent Robots
and Systems, pp.1021-1026, 1998.
[25] X.J. Liu, J.S. Wang and F. Gao, On the optimum design of planar 3-DOF
parallel manipulators with respect to the workspace, in Proc. IEEE Int. Conf.
on Robotics and Automation, pp.4122-4127, 2000.
172
[26] Z. Wang, Z. Wang, W. Liu and Y. Lei, A study on workspace, boundary

workspace analysis and workpiece positioning for parallel machine tools,
Mechanism and Machine Theory, Vol.36, pp.605-622, 2001.
[27] C. Gosselin and J. Angeles, A global performance index for the kinematics
optimization of robotic manipulators, Journal of Mechanical Design, Vol.113,
pp.220-226, 1991.
[28] L.W. Tsai and S. Joshi, Kinematics and optimization of a spatial 3-UPU
parallel manipulator, Journal of Mechanical Design, Vol.122, pp.439-446,
2000.
[29] S. Leguay-Durand and C. Reboulet, Optimal design of a redundant spherical

parallel manipulator, Robotica, Vol.15, pp.399-405, 1997.
[30] M. Gallant and R. Boudrea, The synthesis of planar parallel manipulators with
prismatic joints for an optimal, singularity-free workspace, Journal of Robotic
System, Vol.19, No.1, pp.13-24, 2002.
[31] J.A. Carretero, R.P. Podhorodeski, M.A. Nahon and C.M. Gosselin,
Kinematics analysis and optimization of a new three degree-of-freedom
spatial parallel manipulator, Journal of Mechanical Design, Vol.122, pp.17-24,
2000.
[32] S. Bhattachaya, H. Hatwal and A. Ghosh, On the optimum design of stewart

platform type parallel manipulators, Robotica, Vol.13, pp.133-140, 1995.
[33] J. Ryu and J. Cha, Optimal architecture design of parallel manipulators for
best accuracy, in Proc. IEEE/RSJ Int. Conf. on Intelligent Robots and Systems,
pp.1281-1286, 2001.
[34] K. Miller and R. Clavel, The Lagange-based model of Delta-4 robot

dynamics, Robotersysteme, Vol. 8, pp. 49-54, 1992.
173
[35] K.M. Lee and D.K. Shah, Dynamic analysis of a three-degree-of-freedom inparallel actuated manipulator, IEEE Journal of Robotics and Automation, Vol.
4, No.3, pp.361-367, 1988.
[36] K. Miller, Experimental verification of modeling of DELTA robot dynamics

by direct application of Hamiltons principle, in Proc. IEEE Int. Conf. on
[37] K. Miller, Optimal design and modeling of spatial parallel manipulators, The
Int. Journal of Robotics Research, Vol. 23, No.2, pp.127-140, 2004.
[38] L.W. Tsai, Solving the inverse dynamics of parallel manipulators by the
principle of virtual work, DETC98/MECH-5865, Proc. ASME Design
Engineering Technical Conference, 1998.
[39] J. Wang and C. Gosselin, A New Approach for the Dynamic Analysis of
Parallel Manipulators, Multibody System Dynamics, Vol. 2, pp. 317-334, 1998.
[40] B. Dasgupta and T.S. Mruthyunjaya, Closed-form dynamic equations of the

general Stewart platform through the Newton-Euler approach, Mechanism and
Machine Theory, Vol. 33, No.7, pp.993-1012, 1998.
[41] B. Dasgupta and P. Choudhury, A general strategy based on the Newton-Euler

approach for the dynamic formulation of parallel manipulators, Mechanism
and Machine Theory, Vol. 34, pp.801-824, 1999.
[42] C.D. Lee, D.A. Lawrence and L.Y. Pao, Dynamic modeling and parameter
identification of a parallel haptic interface, in Proc. IEEE Virtual Reality
Conference, 2002.
[43] K. Liu, M. Fitzgerald, D.M. Dawson and F.L. Lewis, Modeling and control of
a stewart platform manipulator, Control of Systems with Inexact Dynamic
Models, pp.83-89, ASME, 1991.
174
[44] A. Codourey, Dynamic Modeling of Parallel Robots for Computed-torque

Control Implementation, Int. J. of Robotics Research, Vol. 17, No. 12, pp.
1325-1336, 1998.
[45] M. Honegger, A. Codourey and E. Burdet, Adaptive Control of the Hexaglide,

a 6 dof Parallel Manipulator, in Proc. IEEE Int. Conf. on Robotics and
Automation, pp. 543-548, 1997.
[46] N. Kim and C. Lee, High Speed Tracking Control of Stewart Platform
Manipulator via Enhanced Sliding Mode Control, in Proc. IEEE Int. Conf. on
Robotics and Automation, pp. 2716-2721, 1998.
[47] N. Kim, C. Lee and P. Chang, Sliding mode control with perturbation
estimation: application to motion control of parallel manipulator, Control
Engineering Practice, Vol. 6, pp. 1321-1330, 1998.
[48] W. Verdonck, J. Swevers and J. Samin, Experimental robot identification:

advantages of combining internal and external measurements and of using
periodic excitation, J. Dynamic Systems, Measurement & Control, Vol.123,
No.4, pp.630-636, 2001.
[49] K. Miller, Experimental verification of modelling of DELTA robot dynamics

by direct application of Hamiltons principle, in Proc. IEEE Int. Conf. on
[50] M.M. Olsen and H.G. Petersen, A new method for estimating parameters of a
dynamic robot model, IEEE Trans. Robotics & Automation, Vol.17, No.1,
pp.95-100, 2001.
[51] H.G. Sage, M.F. de Mathelin and E. Ostertag, Robust control of robot
manipulators: a survey, Int. J. of Control, Vol.72, No.16, pp.1498-1522, 1999.
175
[52] P. Chiacchio, F. Pierrot, L. Sciavicco and B. Siciliano, Robust design of

independent joint controllers with experimentation on a high-speed parallel
robot, IEEE Trans. Industrial Electronics, Vol.40, No.4, pp.393-403, 1993.
[53] G.W. Lee and F.T. Cheng, Robust control of manipulators using the computed
torque plus H compensation method, IEE Proc. Control Theory Appl.,
Vol.143, No.1, pp.64-72, 1996.
[54] F. Lin and R.D. Brandt, An optimal control approach to robust control of
robot manipulators, IEEE Trans. Robotics and Automation, Vol. 14, No.1,
pp.69-77, 1998.
[55] D.H. Kim, J.Y. Kang and K.I. Lee, Robust tracking control design for a 6
DOF Parallel Manipulator, J. Robotic Systems, Vol.17, No.10, pp.527-547,
2000.
[56] M. Karkoub, G. Balas, K. Tamma and M. Donath, Robust control of flexible

manipulators via -synthesis, Control Engineering Practice, Vol.8, pp.725734, 2000.
[57] H.S. Choi, Robust control of robot manipulators with torque saturation using
fuzzy logic, Robotica, Vol.19, pp.631-639, 2001.
[58] R. Caballero, M.A. Armada and T. Akinfiev, Robust cascade controller for
nonlinearity actuated biped robots: experimental evaluation, Int. J. Robotics
Research, Vol.23, No.10-11, pp.1075-1095, 2004.
[59] H.Y. Chuang and Y.C. Chang, Dynamics analysis and learning control for 3PRPS platform, Int. J. of Computer Applications in Technology, Vol.14,
Nos.4/5/6, pp.204-214, 2001.
[60] E. Burdet, L. Rey, A. Codourey, A Trivial Method of Learning Control,
IFAC Symposium of Robot Control, 1997.
176
[61] E. Burdet, L. Rey and A. Codourey, A trivial and efficient learning method for
motion and force control, Engineering Applications of Artificial Intelligence,
Vol.14, pp.487-496, 2001.
[62] N. Umeda, K. Tsuruta and K. Inoki, High-speed positioning of an industrial

robot based on Preview-Learning control, Advanced Robotics, Vol.15, No.3,
pp.307-312, 2001.
[63] S. Gopinath and I.N. Kar, Iterative learning control scheme for manipulators
including actuator dynamics, Mechanism and Machine Theory, Vol.39,
pp.1367-1384, 2004.
[64] J.W.F. Cheung, A.W.K. Chung and N.C. Cheung, Self-tuning control of
brushless servo drive for a high performance tracking manipulator, in Proc.
Int. Conf. Power Electronics and Drive Systems, Vol.1, pp.297-301, 1997.
[65] W. Qin and L. Cai, A frequency domain iterative learning control for low
bandwidth system, in Proc. American Control Conference, pp.1262-1267,
2001.
[66] W. Qin and L. Cai, A fourier series based iterative learning control for
nonlinear uncertain systems, in Proc. IEEE/ASME Int. Conf. on Advanced
Intelligent Mechatronics, pp. 482-487, 2001.
[67] L. Cai and W. Huang, Fourier series based learning control and application to
positioning table, Robotics and Autonomous Systems, Vol.32, pp.89-100, 2000.
[68] S.K. Tso and Y.X. Ma, A discrete learning algorithm for robot motion control
with both position and velocity sensing, in Proc. IEEE/RSJ Int. Workshop on
Intelligent Robots and Systems, pp.633-638, 1991.
177
[69] S.K. Tso and Y.X. Ma, Cartesian-based learning control for robots in discretetime formulation, IEEE Trans. on Systems, Man and Cybernetics, Vol.22,
No.5, pp.1198-1204, 1992.
[70] S.K. Tso and Y.X. Ma, Discrete learning control for robots using partial
model knowledge, in Proc. IEEE Int. Conf. on Robotics and Automation,
pp.2039-2044, 1992.
[71] L.Y.X. Ma, T.S. Low and S.K. Tso, A discrete iterative learning controller for
tracking application, in Proc. Asia-Pacific Workshop on Advances in Motion
Control, pp.206-211, 1993.
[72] D. Chablat and P. Wenger, A New Concept of Modular Parallel Mechanism

for Machining Applications, in Proc. IEEE Int. Conf. on Robotics and
Automation, pp.3965-3970, 2003.
[73] T.C Prentice and B.P. Prescott, U.S. Patent 5,886,494, Positioning System,
March 23, 1999.
[74] L. Sciavicco and B. Siciliano, Modelling and control of robot manipulators.

London: Springer, 2000.
[75] J. Canny, A Computational Approach to Edge Detection, IEEE Transactions
on Pattern Analysis and Machine Intelligence, Vol. 8, No. 6, pp. 679-698, 1986.
[76] F.L. Lewis, D.M. Dawson and C.T. Abdallah, Robot Manipulator Control
Theory and Practice, Second Edition. New York: Marcel Dekker, 2004.
[77] L.W. Tsai, Robot Analysis: The Mechanics of Serial and Parallel Manipulators.
New York: Wiley, 1999.
[78] Solid Edge V.14, UGS Corp., 2003.
178
[79] S.W. Tam, Critical Design of the Dual AC Servo Motor Driver. ASM
Assembly Automation Ltd. Internal Report, 2000.
[80] K. Zhou and J.C. Doyle, Essentials of Robust Control. Upper Saddle River,
N.J.: Prentice Hall, 1998.
[81] W. Huang, L. Cai and X. Tang, Adaptive repetitive output feedback control
for friction and backlash compensation of a positioning table, Proc. 37th IEEE
Conf. on Decision & Control, pp.1250-1251, 1998.
[82] F-206 HexAlign, Physik Instrumente GmbH & Co., 2004.
[83] BS 4500-5:1988 ISO limits and fits. Specification for system of cone tolerances
for conical workspace from C = 1:3 to 1:500 and lengths from 6mm to 630mm,
1998-11-30, The British Standard Institution.
[84] C. Gosselin and J. Angeles, The Optimum Kinematic Design of a Planar

Three-Degree-of-Freedom Parallel Manipulator, ASME J. Mech. Transm.
Autom. Des., Vol. 110, pp.202-207, 1988.
[85] H.R. Mohammadi, P.J. Daniali, P.J. Zsombor-Murray and J. Angeles, The
Kinematics of 3-DOF Planar and Spherical Double-Triangular Parallel
Manipulators, Computational Kinematics, pp.153-164, 1993.
179

FullText PDF

Transféré par

Informations du document

Description originale:

Titre original

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

FullText PDF

Transféré par

Droits d'auteur :

Formats disponibles

Title

Kinematics, dynamics and control of high precision parallel

Cheung, Wing-fung, Jacob.; .

The author retains all proprietary rights, (such as patent rights)

Abstracts of thesis entitled

Kinematics, Dynamics and Control of High Precision Parallel

Cheung Wing Fung Jacob

A dynamic model is essential for motion analysis, capability estimation and

Kinematics, Dynamics and Control of High Precision Parallel

Cheung Wing Fung Jacob

A thesis submitted in partial fulfillment of the requirements for

This thesis is dedicated to my parents and my wife

1.2. Literature survey

1.3. Research Objective and Scope

1.5. Thesis Organization

2.2. Parallel Manipulator Development

2.3. Manipulator Structure

2.4. Kinematics of the Planar Parallel Manipulator

2.4.1. Forward Kinematics

2.4.2. Inverse Kinematics

2.5. Kinematics of the 4-DOF Parallel Manipulator

2.5.1. Forward Kinematics

2.5.2. Inverse Kinematics

3. Jacobian, Singularity and Workspace Analysis

3.2. Derivation of Jacobian Matrix

3.3. Singularity Analysis

3.4. Determination of the Workspace Envelope

3.4.1. 3D Delaunay Tessellation Method

3.4.2. Monte-Carlo Simulation Method

4.2. Derivation of Error Manipulation Matrix

4.3. Design of Objective Functions

4.3.1. Minimizing the Integral of EM

over the Operational

over the Operational 53

5. Dynamic Modelling of the Parallel Manipulator

5.2. Dynamic Modelling using Recursive Newton-Euler Method

5.3. Dynamic Modelling of the Planar Mechanism

5.4. Dynamic Modelling of the Vertical Mechanism

5.5. Modelling of the Belt Mechanism

5.6. Dynamic Simulations

5.8.1. Expressions for matrices in the dynamic model (5.11)

5.8.2. Expressions for Matrices in the dynamic model (5.20)

6.2. Mechanical Design of the Manipulator Prototype

6.3. Electronic Design of the Manipulator Prototype

6.4. Experimental Verification of the Dynamic Model

7. Control Design of the Planar Parallel Manipulator

7.2. Computed-torque Control Design

7.3. Robust Control Design

7.4. Robust Learning Control Design

7.5. Experimental Results

8. Control Design of the 4-DOF Parallel Manipulator

8.2. Computed-torque Control Design

8.3. Robust Control Design

8.4. Robust Learning Control Design

8.5. Experimental Results

9. Conclusion and Future Research Work

9.2. Future Research Work

d1, d3, d5, d7

Position of the linear actuators 1, 3, 5, and 7

l2, l4, l6, l8, l

Length of the linkages of the planar triangular mechanisms

Length of the vertical linkages of the 4-DOF parallel